Add files using upload-large-folder tool

FILE_MANIFEST.txt

@@ -117,6 +117,67 @@
 ./artifacts/outputs/interaction_debug/ablation_short_history_actionhist/reveal_benchmark.json
 ./artifacts/outputs/interaction_debug/ablation_short_history_actionhist/reveal_benchmark.md
 ./artifacts/outputs/interaction_debug/chunk_debug_trace.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_dual_push_buttons/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_dual_push_buttons/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item_easy/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item_easy/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_ball/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_ball/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_tray/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_tray/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_laptop/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_laptop/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_plate/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_plate/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_push_box/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_push_box/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_bottle_in_fridge/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_bottle_in_fridge/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_item_in_drawer/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_item_in_drawer/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_straighten_rope/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_straighten_rope/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_sweep_to_dustpan/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_sweep_to_dustpan/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_take_tray_out_of_oven/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_take_tray_out_of_oven/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_dual_push_buttons/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_dual_push_buttons/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item_easy/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item_easy/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_ball/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_ball/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_tray/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_tray/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_laptop/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_laptop/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_plate/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_plate/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_push_box/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_push_box/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_bottle_in_fridge/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_bottle_in_fridge/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_item_in_drawer/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_item_in_drawer/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_straighten_rope/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_straighten_rope/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_sweep_to_dustpan/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_sweep_to_dustpan/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_take_tray_out_of_oven/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_take_tray_out_of_oven/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.md
+./artifacts/outputs/interaction_debug/peract2_bimanual_push_box_launch_smoke.json
+./artifacts/outputs/interaction_debug/peract2_import_smoke.json
+./artifacts/outputs/interaction_debug/peract2_import_train_command.txt
+./artifacts/outputs/interaction_debug/peract2_smoke_push_box_plan/rollout_eval.json
+./artifacts/outputs/interaction_debug/peract2_smoke_push_box_plan/rollout_eval.md
 ./artifacts/outputs/interaction_debug/proxy_interaction_state_actionhist/checkpoint_best.pt
 ./artifacts/outputs/interaction_debug/proxy_interaction_state_actionhist/config_resolved.yaml
 ./artifacts/outputs/interaction_debug/proxy_interaction_state_actionhist/diagnostics/proxy_diagnostics.json
@@ -268,6 +329,7 @@
 ./code/reveal_vla_bimanual/eval/metrics.py
 ./code/reveal_vla_bimanual/eval/report.py
 ./code/reveal_vla_bimanual/eval/run_ablations.py
+./code/reveal_vla_bimanual/eval/run_peract2_task_sweep.py
 ./code/reveal_vla_bimanual/eval/run_proxy_diagnostics.py
 ./code/reveal_vla_bimanual/eval/run_reveal_benchmark.py
 ./code/reveal_vla_bimanual/eval/run_rlbench_rollout_eval.py
@@ -311,8 +373,10 @@
 ./code/reveal_vla_bimanual/sim_rlbench/generate_smoke_dataset.py
 ./code/reveal_vla_bimanual/sim_rlbench/launch_smoke.py
 ./code/reveal_vla_bimanual/sim_rlbench/obs_adapter.py
+./code/reveal_vla_bimanual/sim_rlbench/obs_config.py
 ./code/reveal_vla_bimanual/sim_rlbench/peract2_runner.py
 ./code/reveal_vla_bimanual/sim_rlbench/smoke_test.py
+./code/reveal_vla_bimanual/sim_rlbench/task_resolver.py
 ./code/reveal_vla_bimanual/sim_rlbench/task_splits.py
 ./code/reveal_vla_bimanual/train/__init__.py
 ./code/reveal_vla_bimanual/train/configs/base.yaml
@@ -355,6 +419,7 @@
 ./environment/rlbench_env_explicit.txt
 ./environment/rlbench_env_export.yaml
 ./environment/rlbench_pip_freeze.txt
+./environment/run_peract2_13_rollouts.sh
 ./environment/setup_same_machine.sh
 ./environment/system_packages_same_machine.txt
 ./environment/upstream_revisions.txt

MODEL_INDEX.md

@@ -193,6 +193,27 @@ This file lists the uploaded checkpoints, datasets, and raw report files referen
 - RLBench smoke dataset root
   - `artifacts/data/rlbench_smoke_open_drawer/`
 
+## RLBench PerAct2 13-Task Outputs
+
+- PerAct2 import smoke JSON
+  - `artifacts/outputs/interaction_debug/peract2_import_smoke.json`
+- PerAct2 train command export
+  - `artifacts/outputs/interaction_debug/peract2_import_train_command.txt`
+- PerAct2 bimanual_push_box launch smoke JSON
+  - `artifacts/outputs/interaction_debug/peract2_bimanual_push_box_launch_smoke.json`
+- PerAct2 no-plan 13-task combined rollout JSON
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.json`
+- PerAct2 no-plan 13-task combined rollout Markdown
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.md`
+- PerAct2 planner-enabled 13-task combined rollout JSON
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.json`
+- PerAct2 planner-enabled 13-task combined rollout Markdown
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.md`
+- PerAct2 no-plan per-task rollout root
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/`
+- PerAct2 planner-enabled per-task rollout root
+  - `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/`
+
 ## Smoke Artifacts
 
 - proxy backbone-only smoke checkpoint
@@ -220,3 +241,7 @@ This file lists the uploaded checkpoints, datasets, and raw report files referen
   - `environment/setup_same_machine.sh`
 - same-machine validation helper
   - `environment/validate_same_machine.sh`
+- same-machine PerAct2 rollout helper
+  - `environment/run_peract2_13_rollouts.sh`
+- PerAct2 isolated sweep helper
+  - `code/reveal_vla_bimanual/eval/run_peract2_task_sweep.py`

README.md

@@ -35,6 +35,7 @@ Raw machine outputs are in:
 
 - `environment/setup_same_machine.sh`
 - `environment/validate_same_machine.sh`
+- `environment/run_peract2_13_rollouts.sh`
 - `environment/system_packages_same_machine.txt`
 - `environment/upstream_revisions.txt`
 - `environment/rlbench_env_export.yaml`
@@ -44,6 +45,37 @@ Raw machine outputs are in:
 - `environment/reveal_env_explicit.txt`
 - `environment/reveal_pip_freeze.txt`
 
+## Code Paths Updated In This Bundle
+
+- interaction-state architecture and wiring
+  - `code/reveal_vla_bimanual/models/reveal_head.py`
+  - `code/reveal_vla_bimanual/models/world_model.py`
+  - `code/reveal_vla_bimanual/models/planner.py`
+  - `code/reveal_vla_bimanual/models/action_decoder.py`
+  - `code/reveal_vla_bimanual/models/observation_memory.py`
+  - `code/reveal_vla_bimanual/models/policy.py`
+- proxy training, losses, and evaluator changes
+  - `code/reveal_vla_bimanual/train/losses.py`
+  - `code/reveal_vla_bimanual/train/smoke_checks.py`
+  - `code/reveal_vla_bimanual/eval/run_proxy_diagnostics.py`
+  - `code/reveal_vla_bimanual/eval/run_reveal_benchmark.py`
+- RLBench and PerAct2 adapter changes
+  - `code/reveal_vla_bimanual/sim_rlbench/task_resolver.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/camera_spec.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/obs_config.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/dataset.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/obs_adapter.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/peract2_runner.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/smoke_test.py`
+  - `code/reveal_vla_bimanual/sim_rlbench/launch_smoke.py`
+  - `code/reveal_vla_bimanual/eval/run_rlbench_rollout_eval.py`
+  - `code/reveal_vla_bimanual/eval/run_peract2_task_sweep.py`
+- same-machine bootstrap helpers
+  - `environment/setup_same_machine.sh`
+  - `environment/validate_same_machine.sh`
+  - `environment/run_peract2_13_rollouts.sh`
+  - `environment/upstream_revisions.txt`
+
 ## Raw Training Endpoints
 
 | Run | Checkpoint | Final train total | Final val total | Metrics |
@@ -130,9 +162,9 @@ Source: `artifacts/outputs/interaction/reveal_ablation_interaction_ep1/ablations
 | no_role_tokens | 0.0 | 0.0 | 0.0 | 0.0 | 67.39010111490886 | 0.9835391044616699 | 0.0 | 2.8990648205081624 | 0.7544853488604227 |
 | short_history | 0.0 | 0.0 | 0.0 | 0.0 | 67.46627298990886 | 0.9835391044616699 | 0.0 | 2.900133213897546 | 0.7520630160967509 |
 
-## Raw RLBench Two-Robot Outputs
+## Raw RLBench Outputs
 
-### Env Smoke
+### Legacy Open-Drawer Path
 
 | Output | Raw values | File |
 | --- | --- | --- |
@@ -140,13 +172,44 @@ Source: `artifacts/outputs/interaction/reveal_ablation_interaction_ep1/ablations
 | launch reset smoke | `description=\"hold the drawer with left hand and open the bottom drawer with right hand\"`, `front_rgb_shape=[128,128,3]`, `wrist_rgb_shape=[128,128,3]`, `wrist2_rgb_shape=[128,128,3]`, `right_pose_shape=[7]`, `left_pose_shape=[7]` | `artifacts/outputs/interaction/rlbench_env_smoke/launch_reset_smoke.json` |
 | launch step smoke | `description=\"hold the drawer with left hand and open the bottom drawer with right hand\"`, `reward=0.0`, `done=false`, `front_rgb_shape_after_step=[128,128,3]` | `artifacts/outputs/interaction/rlbench_env_smoke/launch_step_smoke.json` |
 | train command export | raw command text | `artifacts/outputs/interaction/rlbench_env_smoke/import_and_train_command.txt` |
+| offline RLBench smoke summary | `train_dataset.num_samples=1`, `val_dataset.num_samples=1`, `plan_enabled_for_eval=true`, `frozen_modules=[interaction_head, world_model, planner]` | `artifacts/outputs/interaction/rlbench_open_drawer_interaction_smoke/summary.json` |
+| rollout eval | `plan_requested=true`, `plan_applied=true`, `tasks.open_drawer.error=\"A path could not be found because the target is outside of workspace.\"`, `mean_success=0.0` | `artifacts/outputs/interaction/rlbench_open_drawer_rollout_eval/rollout_eval.json` |
 
-### Offline Smoke And Rollout Eval
+### PerAct2 Import And Launch Smoke
 
 | Output | Raw values | File |
 | --- | --- | --- |
-| offline RLBench smoke summary | `train_dataset.num_samples=1`, `val_dataset.num_samples=1`, `plan_enabled_for_eval=true`, `frozen_modules=[interaction_head, world_model, planner]` | `artifacts/outputs/interaction/rlbench_open_drawer_interaction_smoke/summary.json` |
-| rollout eval | `plan_requested=true`, `plan_applied=true`, `tasks.open_drawer.error=\"A path could not be found because the target is outside of workspace.\"`, `mean_success=0.0` | `artifacts/outputs/interaction/rlbench_open_drawer_rollout_eval/rollout_eval.json` |
+| PerAct2 import smoke | `resolved_upstream_root=/workspace/third_party/peract2_bimanual_probe_20260324`, `imports.rlbench=ok`, `imports.pyrep=ok`, `imports.yarr=ok` | `artifacts/outputs/interaction_debug/peract2_import_smoke.json` |
+| PerAct2 train command export | raw command text | `artifacts/outputs/interaction_debug/peract2_import_train_command.txt` |
+| PerAct2 `bimanual_push_box` launch smoke | `task=BimanualPushBox`, `description=\"push the box to the red area\"`, `front_rgb_shape=[128,128,3]`, `wrist_left_rgb_shape=[128,128,3]`, `wrist_right_rgb_shape=[128,128,3]`, `reward=0.0`, `done=false` | `artifacts/outputs/interaction_debug/peract2_bimanual_push_box_launch_smoke.json` |
+
+### PerAct2 13-Task Isolated Rollout Sweeps
+
+| Mode | Plan requested | Plan applied | Mean success | Task count | Error tasks | Combined JSON | Combined Markdown |
+| --- | --- | --- | ---: | ---: | --- | --- | --- |
+| no-plan split | `false` | `false` | 0.0 | 13 | `[]` | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/rollout_eval.md` |
+| planner-enabled split | `true` | `true` | 0.0 | 13 | `[]` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.md` |
+
+Per-task rollout JSON roots:
+
+- `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/`
+- `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/`
+
+| Task | No-plan mean_success | Planner mean_success | No-plan JSON | Planner JSON |
+| --- | ---: | ---: | --- | --- |
+| bimanual_push_box | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_push_box/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_push_box/rollout_eval.json` |
+| bimanual_lift_ball | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_ball/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_ball/rollout_eval.json` |
+| bimanual_dual_push_buttons | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_dual_push_buttons/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_dual_push_buttons/rollout_eval.json` |
+| bimanual_pick_plate | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_plate/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_plate/rollout_eval.json` |
+| bimanual_put_item_in_drawer | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_item_in_drawer/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_item_in_drawer/rollout_eval.json` |
+| bimanual_put_bottle_in_fridge | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_put_bottle_in_fridge/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_bottle_in_fridge/rollout_eval.json` |
+| bimanual_handover_item | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item/rollout_eval.json` |
+| bimanual_pick_laptop | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_pick_laptop/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_laptop/rollout_eval.json` |
+| bimanual_straighten_rope | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_straighten_rope/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_straighten_rope/rollout_eval.json` |
+| bimanual_sweep_to_dustpan | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_sweep_to_dustpan/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_sweep_to_dustpan/rollout_eval.json` |
+| bimanual_lift_tray | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_tray/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_tray/rollout_eval.json` |
+| bimanual_handover_item_easy | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_handover_item_easy/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item_easy/rollout_eval.json` |
+| bimanual_take_tray_out_of_oven | 0.0 | 0.0 | `artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_take_tray_out_of_oven/rollout_eval.json` | `artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_take_tray_out_of_oven/rollout_eval.json` |
 
 ## RLBench Headless Smoke
 
@@ -251,16 +314,19 @@ Use:
 ```bash
 /workspace/VLAarchtests/environment/setup_same_machine.sh
 /workspace/VLAarchtests/environment/validate_same_machine.sh
+/workspace/VLAarchtests/environment/run_peract2_13_rollouts.sh
 ```
 
 That helper:
 
 - links the bundled code to `/workspace/reveal_vla_bimanual` if needed
-- clones the pinned third-party repos into `/workspace/third_party`
+- clones the pinned PerAct2 third-party repos into `/workspace/third_party`
+- optionally clones the legacy two-robot repos when `INSTALL_LEGACY_LAYOUT=1`
 - installs the RLBench env
 - installs the reveal env base
 - installs the RLBench headless X prerequisites
 - starts X on `:99`
+- exposes a one-command 13-task sweep helper through `environment/run_peract2_13_rollouts.sh`
 
 Raw setup metadata is in:
 

artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_ball/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": false,
+  "plan_applied": false,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_lift_ball": {
+      "task_class": "BimanualLiftBall",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_lift_ball/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `False`
+- Plan applied: `False`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_lift_ball`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_push_box/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": false,
+  "plan_applied": false,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_push_box": {
+      "task_class": "BimanualPushBox",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_noplan_split/bimanual_push_box/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `False`
+- Plan applied: `False`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_push_box`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_handover_item": {
+      "task_class": "BimanualHandoverItem",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_handover_item`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item_easy/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_handover_item_easy": {
+      "task_class": "BimanualHandoverItemEasy",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_handover_item_easy/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_handover_item_easy`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_tray/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_lift_tray": {
+      "task_class": "BimanualLiftTray",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_lift_tray/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_lift_tray`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_laptop/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_pick_laptop": {
+      "task_class": "BimanualPickLaptop",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_laptop/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_pick_laptop`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_pick_plate/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_pick_plate`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_bottle_in_fridge/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_put_bottle_in_fridge": {
+      "task_class": "BimanualPutBottleInFridge",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_bottle_in_fridge/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_put_bottle_in_fridge`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_item_in_drawer/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_put_item_in_drawer": {
+      "task_class": "BimanualPutItemInDrawer",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_put_item_in_drawer/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_put_item_in_drawer`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_straighten_rope/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_straighten_rope": {
+      "task_class": "BimanualStraightenRope",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_straighten_rope/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_straighten_rope`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_sweep_to_dustpan/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_sweep_to_dustpan": {
+      "task_class": "BimanualSweepToDustpan",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_sweep_to_dustpan/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_sweep_to_dustpan`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_take_tray_out_of_oven/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_take_tray_out_of_oven": {
+      "task_class": "BimanualTakeTrayOutOfOven",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/bimanual_take_tray_out_of_oven/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_take_tray_out_of_oven`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.json

@@ -0,0 +1,185 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_push_box": {
+      "task_class": "BimanualPushBox",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_lift_ball": {
+      "task_class": "BimanualLiftBall",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_dual_push_buttons": {
+      "task_class": "BimanualDualPushButtons",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_pick_plate": {
+      "task_class": "BimanualPickPlate",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_put_item_in_drawer": {
+      "task_class": "BimanualPutItemInDrawer",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_put_bottle_in_fridge": {
+      "task_class": "BimanualPutBottleInFridge",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_handover_item": {
+      "task_class": "BimanualHandoverItem",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_pick_laptop": {
+      "task_class": "BimanualPickLaptop",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_straighten_rope": {
+      "task_class": "BimanualStraightenRope",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_sweep_to_dustpan": {
+      "task_class": "BimanualSweepToDustpan",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_lift_tray": {
+      "task_class": "BimanualLiftTray",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_handover_item_easy": {
+      "task_class": "BimanualHandoverItemEasy",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    },
+    "bimanual_take_tray_out_of_oven": {
+      "task_class": "BimanualTakeTrayOutOfOven",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_13_rollout_plan_split/rollout_eval.md

@@ -0,0 +1,22 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_push_box`: mean_success=0.000, returns=[0.0]
+- `bimanual_lift_ball`: mean_success=0.000, returns=[0.0]
+- `bimanual_dual_push_buttons`: mean_success=0.000, returns=[0.0]
+- `bimanual_pick_plate`: mean_success=0.000, returns=[0.0]
+- `bimanual_put_item_in_drawer`: mean_success=0.000, returns=[0.0]
+- `bimanual_put_bottle_in_fridge`: mean_success=0.000, returns=[0.0]
+- `bimanual_handover_item`: mean_success=0.000, returns=[0.0]
+- `bimanual_pick_laptop`: mean_success=0.000, returns=[0.0]
+- `bimanual_straighten_rope`: mean_success=0.000, returns=[0.0]
+- `bimanual_sweep_to_dustpan`: mean_success=0.000, returns=[0.0]
+- `bimanual_lift_tray`: mean_success=0.000, returns=[0.0]
+- `bimanual_handover_item_easy`: mean_success=0.000, returns=[0.0]
+- `bimanual_take_tray_out_of_oven`: mean_success=0.000, returns=[0.0]

artifacts/outputs/interaction_debug/peract2_bimanual_push_box_launch_smoke.json

@@ -0,0 +1,35 @@
+{
+  "display": ":99",
+  "headless": true,
+  "task": "BimanualPushBox",
+  "description": "push the box to the red area",
+  "front_rgb_shape": [
+    128,
+    128,
+    3
+  ],
+  "wrist_left_rgb_shape": [
+    128,
+    128,
+    3
+  ],
+  "wrist_right_rgb_shape": [
+    128,
+    128,
+    3
+  ],
+  "right_pose_shape": [
+    7
+  ],
+  "left_pose_shape": [
+    7
+  ],
+  "stepped_mode": "bimanual_noop",
+  "reward": 0.0,
+  "done": false,
+  "front_rgb_shape_after_step": [
+    128,
+    128,
+    3
+  ]
+}

artifacts/outputs/interaction_debug/peract2_import_smoke.json

@@ -0,0 +1,18 @@
+{
+  "camera_names": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "resolution": [
+    224,
+    224
+  ],
+  "global_camera": "front",
+  "resolved_upstream_root": "/workspace/third_party/peract2_bimanual_probe_20260324",
+  "imports": {
+    "rlbench": "ok",
+    "pyrep": "ok",
+    "yarr": "ok"
+  }
+}

artifacts/outputs/interaction_debug/peract2_import_train_command.txt

@@ -0,0 +1 @@
+xvfb-run -a -s '-screen 0 1280x1024x24' /workspace/envs/rlbench/bin/python train.py method=PERACT_BC rlbench.task_name=multi_3cam 'rlbench.tasks=[bimanual_push_box,bimanual_lift_ball,bimanual_dual_push_buttons,bimanual_pick_plate,bimanual_put_item_in_drawer,bimanual_put_bottle_in_fridge,bimanual_handover_item,bimanual_pick_laptop,bimanual_straighten_rope,bimanual_sweep_to_dustpan,bimanual_lift_tray,bimanual_handover_item_easy,bimanual_take_tray_out_of_oven]' rlbench.demos=100 rlbench.demo_path=/workspace/data/rlbench2 replay.path=/workspace/replays/rlbench2 framework.logdir=/workspace/logs/rlbench2 framework.training_iterations=40000 framework.gpu=0 framework.env_gpu=0 framework.start_seed=0 ddp.num_devices=1 'rlbench.cameras=[front,wrist_left,wrist_right]' 'rlbench.camera_resolution=[224,224]'

artifacts/outputs/interaction_debug/peract2_smoke_push_box_plan/rollout_eval.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint": "/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt",
+  "plan_requested": true,
+  "plan_applied": true,
+  "planner_mode": "trainable",
+  "support_mode_conditioning": true,
+  "episodes_per_task": 1,
+  "episode_length": 20,
+  "resolution": 224,
+  "cameras": [
+    "front",
+    "wrist_left",
+    "wrist_right"
+  ],
+  "tasks": {
+    "bimanual_push_box": {
+      "task_class": "BimanualPushBox",
+      "successes": [
+        0.0
+      ],
+      "returns": [
+        0.0
+      ],
+      "path_recoveries": 0,
+      "noop_fallbacks": 0,
+      "mean_success": 0.0,
+      "mean_return": 0.0
+    }
+  },
+  "mean_success": 0.0
+}

artifacts/outputs/interaction_debug/peract2_smoke_push_box_plan/rollout_eval.md

@@ -0,0 +1,11 @@
+# RLBench Rollout Eval
+
+- Checkpoint: `/workspace/VLAarchtests/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt`
+- Plan requested: `True`
+- Plan applied: `True`
+- Support-mode conditioning: `True`
+- Mean success: `0.000`
+
+## Per-task
+
+- `bimanual_push_box`: mean_success=0.000, returns=[0.0]

code/reveal_vla_bimanual/eval/run_peract2_task_sweep.py

@@ -0,0 +1,210 @@
+from __future__ import annotations
+
+import argparse
+import json
+import subprocess
+import sys
+from pathlib import Path
+from typing import Any
+
+from sim_rlbench.task_splits import PERACT2_BIMANUAL_TASKS
+
+
+def _run_task(
+    project_root: Path,
+    checkpoint: Path,
+    output_dir: Path,
+    task_name: str,
+    *,
+    plan: bool,
+    episodes_per_task: int,
+    episode_length: int,
+    resolution: int,
+    device: str,
+    headless: bool,
+    chunk_commit_steps: int,
+    allow_unsupervised_planning: bool,
+    disable_support_mode_conditioning: bool,
+) -> dict[str, Any]:
+    task_dir = output_dir / task_name
+    task_dir.mkdir(parents=True, exist_ok=True)
+    command = [
+        sys.executable,
+        "-m",
+        "eval.run_rlbench_rollout_eval",
+        "--checkpoint",
+        str(checkpoint),
+        "--output-dir",
+        str(task_dir),
+        "--tasks",
+        task_name,
+        "--episodes-per-task",
+        str(episodes_per_task),
+        "--episode-length",
+        str(episode_length),
+        "--resolution",
+        str(resolution),
+        "--device",
+        device,
+        "--chunk-commit-steps",
+        str(chunk_commit_steps),
+    ]
+    if headless:
+        command.append("--headless")
+    if plan:
+        command.append("--plan")
+    if allow_unsupervised_planning:
+        command.append("--allow-unsupervised-planning")
+    if disable_support_mode_conditioning:
+        command.append("--disable-support-mode-conditioning")
+
+    completed = subprocess.run(
+        command,
+        cwd=project_root,
+        text=True,
+        capture_output=True,
+        check=False,
+    )
+    (task_dir / "command.txt").write_text(" ".join(command) + "\n", encoding="utf-8")
+    (task_dir / "stdout.txt").write_text(completed.stdout, encoding="utf-8")
+    (task_dir / "stderr.txt").write_text(completed.stderr, encoding="utf-8")
+
+    rollout_path = task_dir / "rollout_eval.json"
+    if rollout_path.exists():
+        payload = json.loads(rollout_path.read_text(encoding="utf-8"))
+        task_payload = payload.get("tasks", {}).get(task_name, {})
+    else:
+        task_payload = {}
+
+    if completed.returncode != 0 and "error" not in task_payload:
+        task_payload["error"] = f"subprocess_exit_{completed.returncode}"
+    if "mean_success" not in task_payload:
+        task_payload["mean_success"] = 0.0
+    if "mean_return" not in task_payload:
+        task_payload["mean_return"] = 0.0
+    task_payload["subprocess_returncode"] = completed.returncode
+    task_payload["rollout_path"] = str(rollout_path)
+    return task_payload
+
+
+def _mode_name(plan: bool) -> str:
+    return "plan" if plan else "noplan"
+
+
+def _mode_output_dir(output_root: Path, run_name_prefix: str, plan: bool) -> Path:
+    return output_root / f"{run_name_prefix}_{_mode_name(plan)}_split"
+
+
+def _write_summary_markdown(path: Path, payload: dict[str, Any]) -> None:
+    lines = [
+        "# PerAct2 13-Task Rollout Sweep",
+        "",
+        f"- Checkpoint: `{payload['checkpoint']}`",
+        f"- Plan requested: `{payload['plan_requested']}`",
+        f"- Plan applied: `{payload['plan_applied']}`",
+        f"- Episodes per task: `{payload['episodes_per_task']}`",
+        f"- Episode length: `{payload['episode_length']}`",
+        f"- Resolution: `{payload['resolution']}`",
+        f"- Task count: `{payload['task_count']}`",
+        f"- Error tasks: `{payload['error_tasks']}`",
+        f"- Mean success: `{payload['mean_success']:.3f}`",
+        "",
+        "## Per-task",
+        "",
+    ]
+    for task_name, task_payload in payload["tasks"].items():
+        if "error" in task_payload:
+            lines.append(
+                f"- `{task_name}`: mean_success={task_payload['mean_success']:.3f}, "
+                f"mean_return={task_payload['mean_return']:.3f}, "
+                f"error={task_payload['error']}, "
+                f"subprocess_returncode={task_payload['subprocess_returncode']}"
+            )
+            continue
+        lines.append(
+            f"- `{task_name}`: mean_success={task_payload['mean_success']:.3f}, "
+            f"mean_return={task_payload['mean_return']:.3f}, "
+            f"path_recoveries={task_payload.get('path_recoveries')}, "
+            f"noop_fallbacks={task_payload.get('noop_fallbacks')}"
+        )
+    path.write_text("\n".join(lines) + "\n", encoding="utf-8")
+
+
+def _run_mode(args: argparse.Namespace, plan: bool) -> Path:
+    project_root = Path(__file__).resolve().parents[1]
+    checkpoint = Path(args.checkpoint).resolve()
+    output_dir = _mode_output_dir(Path(args.output_root).resolve(), args.run_name_prefix, plan)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    summary: dict[str, Any] = {
+        "checkpoint": str(checkpoint),
+        "plan_requested": plan,
+        "plan_applied": plan,
+        "episodes_per_task": args.episodes_per_task,
+        "episode_length": args.episode_length,
+        "resolution": args.resolution,
+        "device": args.device,
+        "tasks": {},
+        "subprocess_mode": "isolated_per_task",
+    }
+
+    tasks = tuple(args.tasks) if args.tasks else PERACT2_BIMANUAL_TASKS
+    for task_name in tasks:
+        print(f"[peract2-sweep] running task={task_name} plan={plan}", flush=True)
+        summary["tasks"][task_name] = _run_task(
+            project_root,
+            checkpoint,
+            output_dir,
+            task_name,
+            plan=plan,
+            episodes_per_task=args.episodes_per_task,
+            episode_length=args.episode_length,
+            resolution=args.resolution,
+            device=args.device,
+            headless=args.headless,
+            chunk_commit_steps=args.chunk_commit_steps,
+            allow_unsupervised_planning=args.allow_unsupervised_planning,
+            disable_support_mode_conditioning=args.disable_support_mode_conditioning,
+        )
+
+    task_scores = [float(task_payload["mean_success"]) for task_payload in summary["tasks"].values()]
+    summary["task_count"] = len(summary["tasks"])
+    summary["error_tasks"] = sorted(
+        task_name for task_name, task_payload in summary["tasks"].items() if "error" in task_payload
+    )
+    summary["mean_success"] = float(sum(task_scores) / len(task_scores)) if task_scores else 0.0
+
+    summary_path = output_dir / "rollout_eval.json"
+    summary_path.write_text(json.dumps(summary, indent=2), encoding="utf-8")
+    _write_summary_markdown(output_dir / "rollout_eval.md", summary)
+    return summary_path
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--checkpoint", required=True)
+    parser.add_argument("--output-root", required=True)
+    parser.add_argument("--run-name-prefix", default="peract2_13_rollout")
+    parser.add_argument("--tasks", nargs="*", default=list(PERACT2_BIMANUAL_TASKS))
+    parser.add_argument("--episodes-per-task", type=int, default=1)
+    parser.add_argument("--episode-length", type=int, default=20)
+    parser.add_argument("--resolution", type=int, default=224)
+    parser.add_argument("--device", default="cuda")
+    parser.add_argument("--headless", action="store_true", default=True)
+    parser.add_argument("--chunk-commit-steps", type=int, default=4)
+    parser.add_argument("--allow-unsupervised-planning", action="store_true")
+    parser.add_argument("--disable-support-mode-conditioning", action="store_true")
+    parser.add_argument("--skip-noplan", action="store_true")
+    parser.add_argument("--skip-plan", action="store_true")
+    args = parser.parse_args()
+
+    generated = []
+    if not args.skip_noplan:
+        generated.append(_run_mode(args, plan=False))
+    if not args.skip_plan:
+        generated.append(_run_mode(args, plan=True))
+    print(json.dumps({"generated": [str(path) for path in generated]}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

code/reveal_vla_bimanual/eval/run_rlbench_rollout_eval.py

@@ -7,12 +7,10 @@ from typing import Any, Sequence
 
 import numpy as np
 import torch
-from helpers.utils import create_obs_config
-from rlbench.action_modes.action_mode import MoveArmThenGripper2Robots
-from rlbench.action_modes.arm_action_modes import EndEffectorPoseViaPlanning2Robots
-from rlbench.action_modes.gripper_action_modes import Discrete2Robots
-from rlbench.backend.utils import task_file_to_task_class
-from rlbench.environments_two_robots import Environment2Robots
+from rlbench.action_modes.action_mode import BimanualMoveArmThenGripper
+from rlbench.action_modes.arm_action_modes import BimanualEndEffectorPoseViaPlanning
+from rlbench.action_modes.gripper_action_modes import BimanualDiscrete
+from rlbench.environment import Environment
 
 from models.action_decoder import ChunkDecoderConfig
 from models.backbones import FrozenVLBackboneConfig
@@ -24,10 +22,12 @@ from models.reveal_head import RevealHeadConfig
 from models.world_model import RevealWMConfig
 from sim_rlbench.camera_spec import default_three_camera_spec
 from sim_rlbench.dataset import (
+    absolute_action_from_delta,
     bimanual_proprio_from_obs,
-    single_arm_absolute_action_from_delta,
     stack_live_rgb_obs,
 )
+from sim_rlbench.obs_config import build_obs_config
+from sim_rlbench.task_resolver import resolve_task_class
 from train.trainer import TrainerConfig, build_policy, planner_enabled, policy_supports_planning
 
 
@@ -52,66 +52,51 @@ def _episode_language_goal(descriptions: Sequence[str]) -> str:
     return str(descriptions[0]) if descriptions else ""
 
 
-def _noop_arm_action(obs: Any, arm_name: str) -> np.ndarray:
-    if arm_name == "right":
-        pose = np.asarray(obs.gripper_right_pose, dtype=np.float32)
-        gripper_open = float(obs.gripper_right_open)
-    elif arm_name == "left":
-        pose = np.asarray(obs.gripper_left_pose, dtype=np.float32)
-        gripper_open = float(obs.gripper_left_open)
-    else:  # pragma: no cover - defensive guard
-        raise ValueError(f"Unsupported arm: {arm_name}")
-    return np.concatenate([pose, np.array([gripper_open, 1.0], dtype=np.float32)], axis=0)
+def _noop_bimanual_action(obs: Any) -> np.ndarray:
+    right_obs = getattr(obs, "right", None)
+    left_obs = getattr(obs, "left", None)
+    right = np.concatenate(
+        [
+            np.asarray(right_obs.gripper_pose, dtype=np.float32),
+            np.array([float(right_obs.gripper_open), 1.0], dtype=np.float32),
+        ],
+        axis=0,
+    )
+    left = np.concatenate(
+        [
+            np.asarray(left_obs.gripper_pose, dtype=np.float32),
+            np.array([float(left_obs.gripper_open), 1.0], dtype=np.float32),
+        ],
+        axis=0,
+    )
+    return np.concatenate([right, left], axis=0)
 
 
-def _scaled_single_arm_delta(delta_action: np.ndarray, arm_name: str, scale: float) -> np.ndarray:
+def _scaled_bimanual_delta(delta_action: np.ndarray, scale: float) -> np.ndarray:
     scaled = np.asarray(delta_action, dtype=np.float32).copy()
-    arm_index = {"right": 0, "left": 1}[arm_name]
-    offset = arm_index * 7
-    scaled[offset : offset + 6] *= float(scale)
+    scaled[0:6] *= float(scale)
+    scaled[7:13] *= float(scale)
     return scaled
 
 
 def _step_bimanual_chunk(task: Any, obs: Any, delta_action: np.ndarray) -> tuple[Any, float, bool, int, int]:
-    total_reward = 0.0
-    done = False
-    next_obs = obs
-    recovered_steps = 0
-    noop_fallbacks = 0
-    for arm_name in ("right", "left"):
-        reward = 0.0
-        last_error: Exception | None = None
-        stepped = False
-        for scale in (1.0, 0.5, 0.25, 0.1):
-            try:
-                env_action = single_arm_absolute_action_from_delta(
-                    next_obs,
-                    _scaled_single_arm_delta(delta_action, arm_name, scale),
-                    arm_name,
-                    ignore_collisions=True,
-                )
-                next_obs, reward, done = task.step(env_action, arm_name)
-                if scale < 1.0:
-                    recovered_steps += 1
-                stepped = True
-                break
-            except Exception as exc:  # pragma: no cover - live RLBench failure path
-                last_error = exc
-        if not stepped:
-            try:
-                next_obs, reward, done = task.step(_noop_arm_action(next_obs, arm_name), arm_name)
-                noop_fallbacks += 1
-                stepped = True
-            except Exception as exc:  # pragma: no cover - live RLBench failure path
-                last_error = exc
-        if not stepped:
-            if last_error is not None:
-                raise last_error
-            raise RuntimeError(f"Failed to step arm '{arm_name}' for unknown reasons.")
-        total_reward += float(reward)
-        if reward >= 1.0 or done:
-            break
-    return next_obs, total_reward, done, recovered_steps, noop_fallbacks
+    last_error: Exception | None = None
+    for scale in (1.0, 0.5, 0.25, 0.1):
+        try:
+            env_action = absolute_action_from_delta(obs, _scaled_bimanual_delta(delta_action, scale), ignore_collisions=True)
+            next_obs, reward, done = task.step(env_action)
+            recovered_steps = 1 if scale < 1.0 else 0
+            return next_obs, float(reward), bool(done), recovered_steps, 0
+        except Exception as exc:  # pragma: no cover - live RLBench failure path
+            last_error = exc
+    try:
+        next_obs, reward, done = task.step(_noop_bimanual_action(obs))
+        return next_obs, float(reward), bool(done), 0, 1
+    except Exception as exc:  # pragma: no cover - live RLBench failure path
+        last_error = exc
+    if last_error is not None:
+        raise last_error
+    raise RuntimeError("Failed to step bimanual action for unknown reasons.")
 
 
 def main() -> None:
@@ -179,24 +164,19 @@ def main() -> None:
     for task_name in args.tasks:
         task_successes: list[float] = []
         task_returns: list[float] = []
-        env: Environment2Robots | None = None
+        env: Environment | None = None
         try:
-            task_class = task_file_to_task_class(task_name)
-            obs_config = create_obs_config(
-                list(camera_spec.upstream_cameras),
-                [args.resolution, args.resolution],
-                "PERACT_BC",
-            )
-            action_mode = MoveArmThenGripper2Robots(
-                EndEffectorPoseViaPlanning2Robots(absolute_mode=True, frame="world", collision_checking=False),
-                Discrete2Robots(),
+            task_class = resolve_task_class(task_name)
+            obs_config = build_obs_config(list(camera_spec.upstream_cameras), args.resolution)
+            action_mode = BimanualMoveArmThenGripper(
+                BimanualEndEffectorPoseViaPlanning(absolute_mode=True, frame="world", collision_checking=False),
+                BimanualDiscrete(),
             )
-            env = Environment2Robots(
+            env = Environment(
                 action_mode=action_mode,
                 obs_config=obs_config,
                 headless=args.headless,
-                robot_setup="panda",
-                task_name=task_class.__name__,
+                robot_setup="dual_panda",
             )
             env.launch()
             task = env.get_task(task_class)

code/reveal_vla_bimanual/models/backbones.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 import math
+from pathlib import Path
 from typing import Sequence
 
 import torch
@@ -48,12 +49,38 @@ class FrozenVLBackbone(nn.Module):
         else:
             from transformers import AutoTokenizer, CLIPModel
 
-            clip_model = CLIPModel.from_pretrained(config.model_name, use_safetensors=True)
+            local_model_source: str | None = None
+            if config.model_name == "openai/clip-vit-base-patch32":
+                cache_root = Path("/workspace/.cache/huggingface/hub/models--openai--clip-vit-base-patch32")
+                ref_path = cache_root / "refs" / "main"
+                if ref_path.exists():
+                    snapshot_id = ref_path.read_text(encoding="utf-8").strip()
+                    snapshot_dir = cache_root / "snapshots" / snapshot_id
+                    if (snapshot_dir / "config.json").exists():
+                        local_model_source = str(snapshot_dir)
+            clip_model = None
+            if local_model_source is not None:
+                try:
+                    clip_model = CLIPModel.from_pretrained(
+                        local_model_source,
+                        use_safetensors=True,
+                        local_files_only=True,
+                    )
+                except OSError:
+                    clip_model = None
+            if clip_model is None:
+                clip_model = CLIPModel.from_pretrained(config.model_name, use_safetensors=True)
             self.vision_model = clip_model.vision_model
             self.text_model = clip_model.text_model
             self.visual_projection = clip_model.visual_projection
             self.text_projection = clip_model.text_projection
-            self.tokenizer = AutoTokenizer.from_pretrained(config.model_name)
+            if local_model_source is not None:
+                try:
+                    self.tokenizer = AutoTokenizer.from_pretrained(local_model_source, local_files_only=True)
+                except OSError:
+                    self.tokenizer = AutoTokenizer.from_pretrained(config.model_name)
+            else:
+                self.tokenizer = AutoTokenizer.from_pretrained(config.model_name)
             self.hidden_dim = clip_model.config.projection_dim
             if config.gradient_checkpointing:
                 if hasattr(self.vision_model, "gradient_checkpointing_enable"):

code/reveal_vla_bimanual/scripts/setup_env_a_rlbench.sh

@@ -58,10 +58,21 @@ run_in_env python -m pip install --force-reinstall --no-deps numpy==1.26.4 Pillo
 run_in_env python -m pip install --no-deps moviepy==2.2.1 timeout-decorator==0.5.0 opencv-python==4.10.0.84 pyquaternion==0.9.9 click-prompt==0.5.1
 run_in_env python -m pip install --no-deps poetry-core
 run_in_env python -m pip install gymnasium==1.0.0a2 open3d==0.19.0 segment-anything==1.0 transforms3d==0.4.1 openai==0.28.1
-PERACT_ROOT="${ROOT_DIR}/third_party/peract_bimanual"
-YARR_ROOT="${ROOT_DIR}/third_party/YARR"
-PYREP_ROOT="${ROOT_DIR}/third_party/PyRep"
-RLBENCH_ROOT="${ROOT_DIR}/third_party/RLBench"
+first_existing_dir() {
+  local candidate=""
+  for candidate in "$@"; do
+    if [[ -d "${candidate}" ]]; then
+      printf '%s\n' "${candidate}"
+      return 0
+    fi
+  done
+  printf '%s\n' "$1"
+}
+
+PERACT_ROOT="$(first_existing_dir "${ROOT_DIR}/third_party/peract2_bimanual_probe_20260324" "${ROOT_DIR}/third_party/peract_bimanual")"
+YARR_ROOT="$(first_existing_dir "${ROOT_DIR}/third_party/YARR_peract2_20260324" "${ROOT_DIR}/third_party/YARR")"
+PYREP_ROOT="$(first_existing_dir "${ROOT_DIR}/third_party/PyRep_peract2_20260324" "${ROOT_DIR}/third_party/PyRep")"
+RLBENCH_ROOT="$(first_existing_dir "${ROOT_DIR}/third_party/RLBench_peract2_20260324" "${ROOT_DIR}/third_party/RLBench")"
 if [[ -f "${PERACT_ROOT}/YARR/setup.py" ]]; then
   YARR_ROOT="${PERACT_ROOT}/YARR"
 fi

code/reveal_vla_bimanual/sim_rlbench/camera_spec.py

@@ -4,16 +4,29 @@ from dataclasses import dataclass
 
 CANONICAL_TO_UPSTREAM_CAMERA = {
     "front": "front",
-    "wrist_left": "wrist2",
-    "wrist_right": "wrist",
+    "wrist_left": "wrist_left",
+    "wrist_right": "wrist_right",
+}
+CANONICAL_CAMERA_ALIASES = {
+    "front": ("front",),
+    "wrist_left": ("wrist_left", "wrist2"),
+    "wrist_right": ("wrist_right", "wrist"),
+}
+UPSTREAM_TO_CANONICAL_CAMERA = {
+    alias: camera_name
+    for camera_name, aliases in CANONICAL_CAMERA_ALIASES.items()
+    for alias in aliases
 }
-UPSTREAM_TO_CANONICAL_CAMERA = {value: key for key, value in CANONICAL_TO_UPSTREAM_CAMERA.items()}
 
 
 def canonical_to_upstream_camera(camera_name: str) -> str:
     return CANONICAL_TO_UPSTREAM_CAMERA.get(camera_name, camera_name)
 
 
+def camera_aliases(camera_name: str) -> tuple[str, ...]:
+    return CANONICAL_CAMERA_ALIASES.get(camera_name, (camera_name,))
+
+
 def upstream_to_canonical_camera(camera_name: str) -> str:
     return UPSTREAM_TO_CANONICAL_CAMERA.get(camera_name, camera_name)
 
@@ -33,7 +46,7 @@ class RLBenchThreeCameraSpec:
 
     @property
     def upstream_cameras(self) -> tuple[str, str, str]:
-        return tuple(CANONICAL_TO_UPSTREAM_CAMERA.get(camera, camera) for camera in self.cameras)  # type: ignore[return-value]
+        return tuple(canonical_to_upstream_camera(camera) for camera in self.cameras)  # type: ignore[return-value]
 
     def hydra_overrides(self, prefix: str = "rlbench") -> list[str]:
         camera_list = ",".join(self.upstream_cameras)

code/reveal_vla_bimanual/sim_rlbench/dataset.py

@@ -10,7 +10,7 @@ import torch
 from PIL import Image
 from torch.utils.data import Dataset
 
-from sim_rlbench.camera_spec import canonical_to_upstream_camera
+from sim_rlbench.camera_spec import camera_aliases, canonical_to_upstream_camera
 
 
 THREE_CAMERAS: tuple[str, str, str] = ("front", "wrist_left", "wrist_right")
@@ -27,18 +27,12 @@ def _clip_vector_norm(vector: Sequence[float], max_norm: float) -> np.ndarray:
 
 
 def _camera_value(obs: Any, camera_name: str, suffix: str) -> Any:
-    upstream_name = canonical_to_upstream_camera(camera_name)
-    candidate_keys = [
-        f"{upstream_name}_{suffix}",
-        f"{camera_name}_{suffix}",
-    ]
+    candidate_keys: list[str] = []
+    for alias in camera_aliases(camera_name):
+        candidate_keys.append(f"{alias}_{suffix}")
     if suffix == "point_cloud":
-        candidate_keys.extend(
-            [
-                f"{upstream_name}_pointcloud",
-                f"{camera_name}_pointcloud",
-            ]
-        )
+        for alias in camera_aliases(camera_name):
+            candidate_keys.append(f"{alias}_pointcloud")
     for key in candidate_keys:
         if hasattr(obs, key):
             return getattr(obs, key)
@@ -338,10 +332,16 @@ class RLBenchOfflineChunkDataset(Dataset[dict[str, Any]]):
     def _load_rgb_stack(self, episode_dir: Path, step_index: int) -> torch.Tensor:
         frames: list[np.ndarray] = []
         for camera_name in self.cameras:
-            image_path = episode_dir / f"{camera_name}_rgb" / f"rgb_{step_index:04d}.png"
-            if not image_path.exists():
-                upstream_camera_name = canonical_to_upstream_camera(camera_name)
-                image_path = episode_dir / f"{upstream_camera_name}_rgb" / f"rgb_{step_index:04d}.png"
+            image_path = None
+            for camera_alias in camera_aliases(camera_name):
+                candidate = episode_dir / f"{camera_alias}_rgb" / f"rgb_{step_index:04d}.png"
+                if candidate.exists():
+                    image_path = candidate
+                    break
+            if image_path is None:
+                raise FileNotFoundError(
+                    f"Could not locate RGB frame for camera '{camera_name}' step {step_index} in {episode_dir}"
+                )
             image = Image.open(image_path).convert("RGB")
             if image.size != (self.resolution, self.resolution):
                 image = image.resize((self.resolution, self.resolution), Image.Resampling.BILINEAR)

code/reveal_vla_bimanual/sim_rlbench/generate_smoke_dataset.py

@@ -7,10 +7,9 @@ from pathlib import Path
 
 import numpy as np
 from PIL import Image
-from helpers.utils import create_obs_config
-from rlbench.action_modes.action_mode import MoveArmThenGripper2Robots
-from rlbench.action_modes.arm_action_modes import EndEffectorPoseViaPlanning2Robots
-from rlbench.action_modes.gripper_action_modes import Discrete2Robots
+from rlbench.action_modes.action_mode import BimanualMoveArmThenGripper
+from rlbench.action_modes.arm_action_modes import BimanualEndEffectorPoseViaPlanning
+from rlbench.action_modes.gripper_action_modes import BimanualDiscrete
 from rlbench.backend.const import (
     EPISODE_FOLDER,
     EPISODES_FOLDER,
@@ -19,15 +18,15 @@ from rlbench.backend.const import (
     VARIATION_NUMBER,
     VARIATIONS_ALL_FOLDER,
 )
-from rlbench.backend.utils import task_file_to_task_class
-from rlbench.environments_two_robots import Environment2Robots
+from rlbench.environment import Environment
 
-from sim_rlbench.camera_spec import canonical_to_upstream_camera, default_three_camera_spec
+from sim_rlbench.camera_spec import camera_aliases, default_three_camera_spec
+from sim_rlbench.obs_config import build_obs_config
+from sim_rlbench.task_resolver import resolve_task_class
 
 
 def _camera_payload(obs: object, camera_name: str, suffix: str):
-    upstream_name = canonical_to_upstream_camera(camera_name)
-    for key in (f"{upstream_name}_{suffix}", f"{camera_name}_{suffix}"):
+    for key in [f"{alias}_{suffix}" for alias in camera_aliases(camera_name)]:
         if hasattr(obs, key):
             return getattr(obs, key)
     return None
@@ -70,22 +69,17 @@ def main() -> None:
     args = parser.parse_args()
 
     spec = default_three_camera_spec(args.resolution)
-    obs_config = create_obs_config(
-        list(spec.upstream_cameras),
-        [args.resolution, args.resolution],
-        "PERACT_BC",
-    )
+    obs_config = build_obs_config(list(spec.upstream_cameras), args.resolution)
 
-    task_class = task_file_to_task_class(args.task)
-    env = Environment2Robots(
-        action_mode=MoveArmThenGripper2Robots(
-            EndEffectorPoseViaPlanning2Robots(absolute_mode=True, frame="world", collision_checking=False),
-            Discrete2Robots(),
+    task_class = resolve_task_class(args.task)
+    env = Environment(
+        action_mode=BimanualMoveArmThenGripper(
+            BimanualEndEffectorPoseViaPlanning(absolute_mode=True, frame="world", collision_checking=False),
+            BimanualDiscrete(),
         ),
         obs_config=obs_config,
         headless=True,
-        robot_setup="panda",
-        task_name=task_class.__name__,
+        robot_setup="dual_panda",
     )
     output_root = Path(args.output_root)
     episodes_root = output_root / args.task / VARIATIONS_ALL_FOLDER / EPISODES_FOLDER

code/reveal_vla_bimanual/sim_rlbench/launch_smoke.py

@@ -5,73 +5,88 @@ import json
 import os
 
 import numpy as np
-from helpers.utils import create_obs_config
-from rlbench.action_modes.action_mode import MoveArmThenGripper2Robots
-from rlbench.action_modes.arm_action_modes import EndEffectorPoseViaPlanning2Robots
-from rlbench.action_modes.gripper_action_modes import Discrete2Robots
-from rlbench.backend.utils import task_file_to_task_class
-from rlbench.environments_two_robots import Environment2Robots
+from rlbench.action_modes.action_mode import BimanualMoveArmThenGripper
+from rlbench.action_modes.arm_action_modes import BimanualEndEffectorPoseViaPlanning
+from rlbench.action_modes.gripper_action_modes import BimanualDiscrete
+from rlbench.environment import Environment
 
 from sim_rlbench.camera_spec import default_three_camera_spec
+from sim_rlbench.obs_config import build_obs_config
+from sim_rlbench.task_resolver import resolve_task_class
 
 
-def _noop_arm_action(obs: object, arm_name: str) -> np.ndarray:
-    if arm_name == "right":
-        pose = np.asarray(obs.gripper_right_pose, dtype=np.float32)
-        gripper_open = float(obs.gripper_right_open)
-    elif arm_name == "left":
-        pose = np.asarray(obs.gripper_left_pose, dtype=np.float32)
-        gripper_open = float(obs.gripper_left_open)
-    else:  # pragma: no cover - smoke-only guard
-        raise ValueError(f"Unsupported arm: {arm_name}")
-    return np.concatenate([pose, np.array([gripper_open, 1.0], dtype=np.float32)], axis=0)
+def _camera_shape(obs: object, camera_name: str):
+    perception_data = getattr(obs, "perception_data", None)
+    if isinstance(perception_data, dict):
+        value = perception_data.get(f"{camera_name}_rgb")
+        if value is not None:
+            return list(np.asarray(value).shape)
+    return None
+
+
+def _noop_bimanual_action(obs: object) -> np.ndarray:
+    right_obs = getattr(obs, "right", None)
+    left_obs = getattr(obs, "left", None)
+    right = np.concatenate(
+        [
+            np.asarray(right_obs.gripper_pose, dtype=np.float32),
+            np.array([float(right_obs.gripper_open), 1.0], dtype=np.float32),
+        ],
+        axis=0,
+    )
+    left = np.concatenate(
+        [
+            np.asarray(left_obs.gripper_pose, dtype=np.float32),
+            np.array([float(left_obs.gripper_open), 1.0], dtype=np.float32),
+        ],
+        axis=0,
+    )
+    return np.concatenate([right, left], axis=0)
 
 
 def main() -> None:
     parser = argparse.ArgumentParser()
-    parser.add_argument("--task", default="open_drawer")
+    parser.add_argument("--task", default="bimanual_push_box")
     parser.add_argument("--resolution", type=int, default=224)
     parser.add_argument("--display", default=None)
     parser.add_argument("--headless", action="store_true", default=True)
     parser.add_argument("--visible", action="store_true")
-    parser.add_argument("--arm", choices=("right", "left"), default="right")
     args = parser.parse_args()
 
     headless = args.headless and not args.visible
     spec = default_three_camera_spec(args.resolution)
-    task_cls = task_file_to_task_class(args.task)
-    obs_config = create_obs_config(list(spec.upstream_cameras), [args.resolution, args.resolution], "PERACT_BC")
-    action_mode = MoveArmThenGripper2Robots(
-        EndEffectorPoseViaPlanning2Robots(absolute_mode=True, frame="world", collision_checking=False),
-        Discrete2Robots(),
+    task_cls = resolve_task_class(args.task)
+    obs_config = build_obs_config(list(spec.upstream_cameras), args.resolution)
+    action_mode = BimanualMoveArmThenGripper(
+        BimanualEndEffectorPoseViaPlanning(absolute_mode=True, frame="world", collision_checking=False),
+        BimanualDiscrete(),
     )
-    env = Environment2Robots(
+    env = Environment(
         action_mode=action_mode,
         obs_config=obs_config,
         headless=headless,
-        robot_setup="panda",
-        task_name=task_cls.__name__,
+        robot_setup="dual_panda",
     )
 
     try:
         env.launch()
         task = env.get_task(task_cls)
         descriptions, obs = task.reset()
-        next_obs, reward, done = task.step(_noop_arm_action(obs, args.arm), args.arm)
+        next_obs, reward, done = task.step(_noop_bimanual_action(obs))
         payload = {
             "display": args.display or os.environ.get("DISPLAY"),
             "headless": headless,
             "task": task_cls.__name__,
             "description": descriptions[0] if descriptions else "",
-            "front_rgb_shape": None if obs.front_rgb is None else list(obs.front_rgb.shape),
-            "wrist_rgb_shape": None if obs.wrist_rgb is None else list(obs.wrist_rgb.shape),
-            "wrist2_rgb_shape": None if obs.wrist2_rgb is None else list(obs.wrist2_rgb.shape),
-            "right_pose_shape": None if obs.gripper_right_pose is None else list(obs.gripper_right_pose.shape),
-            "left_pose_shape": None if obs.gripper_left_pose is None else list(obs.gripper_left_pose.shape),
-            "stepped_arm": args.arm,
+            "front_rgb_shape": _camera_shape(obs, "front"),
+            "wrist_left_rgb_shape": _camera_shape(obs, "wrist_left"),
+            "wrist_right_rgb_shape": _camera_shape(obs, "wrist_right"),
+            "right_pose_shape": list(obs.right.gripper_pose.shape),
+            "left_pose_shape": list(obs.left.gripper_pose.shape),
+            "stepped_mode": "bimanual_noop",
             "reward": float(reward),
             "done": bool(done),
-            "front_rgb_shape_after_step": None if next_obs.front_rgb is None else list(next_obs.front_rgb.shape),
+            "front_rgb_shape_after_step": _camera_shape(next_obs, "front"),
         }
         print(json.dumps(payload, indent=2))
     finally:

code/reveal_vla_bimanual/sim_rlbench/obs_adapter.py

@@ -5,7 +5,7 @@ from typing import Any
 
 import numpy as np
 
-from sim_rlbench.camera_spec import RLBenchThreeCameraSpec, canonical_to_upstream_camera
+from sim_rlbench.camera_spec import RLBenchThreeCameraSpec, camera_aliases
 
 
 @dataclass
@@ -31,50 +31,44 @@ class CanonicalBimanualObservation:
 
 
 def _camera_rgb(obs: Any, camera_name: str) -> np.ndarray:
-    upstream_name = canonical_to_upstream_camera(camera_name)
-    for key in (f"{upstream_name}_rgb", f"{camera_name}_rgb"):
+    candidate_keys = [f"{alias}_rgb" for alias in camera_aliases(camera_name)]
+    for key in candidate_keys:
         if hasattr(obs, key):
             return np.asarray(getattr(obs, key), dtype=np.uint8)
     perception_data = getattr(obs, "perception_data", None)
     if isinstance(perception_data, dict):
-        for key in (f"{upstream_name}_rgb", f"{camera_name}_rgb"):
+        for key in candidate_keys:
             if key in perception_data:
                 return np.asarray(perception_data[key], dtype=np.uint8)
     raise KeyError(f"Observation does not expose RGB for camera '{camera_name}'")
 
 
 def _camera_point_cloud(obs: Any, camera_name: str) -> np.ndarray:
-    upstream_name = canonical_to_upstream_camera(camera_name)
-    for key in (
-        f"{upstream_name}_point_cloud",
-        f"{upstream_name}_pointcloud",
-        f"{camera_name}_point_cloud",
-        f"{camera_name}_pointcloud",
-    ):
+    candidate_keys = []
+    for alias in camera_aliases(camera_name):
+        candidate_keys.extend([f"{alias}_point_cloud", f"{alias}_pointcloud"])
+    for key in candidate_keys:
         if hasattr(obs, key):
             return np.asarray(getattr(obs, key), dtype=np.float32)
     perception_data = getattr(obs, "perception_data", None)
     if isinstance(perception_data, dict):
-        for key in (
-            f"{upstream_name}_point_cloud",
-            f"{upstream_name}_pointcloud",
-            f"{camera_name}_point_cloud",
-            f"{camera_name}_pointcloud",
-        ):
+        for key in candidate_keys:
             if key in perception_data:
                 return np.asarray(perception_data[key], dtype=np.float32)
     raise KeyError(f"Observation does not expose point clouds for camera '{camera_name}'")
 
 
 def _camera_misc(obs: Any, camera_name: str, field_name: str) -> np.ndarray:
-    upstream_name = canonical_to_upstream_camera(camera_name)
     misc = getattr(obs, "misc", {})
-    for key in (
-        f"{upstream_name}_camera_{field_name}",
-        f"{camera_name}_camera_{field_name}",
-        f"{upstream_name}_{field_name}",
-        f"{camera_name}_{field_name}",
-    ):
+    candidate_keys = []
+    for alias in camera_aliases(camera_name):
+        candidate_keys.extend(
+            [
+                f"{alias}_camera_{field_name}",
+                f"{alias}_{field_name}",
+            ]
+        )
+    for key in candidate_keys:
         if key in misc:
             return np.asarray(misc[key], dtype=np.float32)
     raise KeyError(f"Observation misc does not expose {field_name} for camera '{camera_name}'")

code/reveal_vla_bimanual/sim_rlbench/obs_config.py

@@ -0,0 +1,40 @@
+from __future__ import annotations
+
+from rlbench.observation_config import CameraConfig, ObservationConfig
+
+
+def build_obs_config(
+    cameras: list[str],
+    resolution: int,
+    *,
+    rgb: bool = True,
+    depth: bool = False,
+    point_cloud: bool = False,
+    mask: bool = False,
+) -> ObservationConfig:
+    camera_configs = {
+        camera: CameraConfig(
+            rgb=rgb,
+            depth=depth,
+            point_cloud=point_cloud,
+            mask=mask,
+            image_size=(resolution, resolution),
+        )
+        for camera in cameras
+    }
+    return ObservationConfig(
+        camera_configs=camera_configs,
+        joint_velocities=True,
+        joint_positions=True,
+        joint_forces=True,
+        gripper_open=True,
+        gripper_pose=True,
+        gripper_matrix=False,
+        gripper_joint_positions=False,
+        gripper_touch_forces=False,
+        wrist_camera_matrix=False,
+        record_gripper_closing=False,
+        task_low_dim_state=True,
+        record_ignore_collisions=True,
+        robot_name="dual_panda",
+    )

code/reveal_vla_bimanual/sim_rlbench/peract2_runner.py

@@ -40,7 +40,7 @@ def resolve_upstream_root(upstream_root: Path) -> Path:
 
 @dataclass
 class BenchmarkRunSpec:
-    upstream_root: Path = Path("/workspace/third_party/peract_bimanual")
+    upstream_root: Path = Path("/workspace/third_party/peract2_bimanual_probe_20260324")
     demo_path: Path = Path("/workspace/data/rlbench2")
     replay_path: Path = Path("/workspace/replays/rlbench2")
     logdir: Path = Path("/workspace/logs/rlbench2")

code/reveal_vla_bimanual/sim_rlbench/smoke_test.py

@@ -12,7 +12,7 @@ from sim_rlbench.peract2_runner import BenchmarkRunSpec, resolve_upstream_root
 def main() -> None:
     parser = argparse.ArgumentParser()
     parser.add_argument("--demo-path", default="/workspace/data/rlbench2")
-    parser.add_argument("--upstream-root", default="/workspace/third_party/peract_bimanual")
+    parser.add_argument("--upstream-root", default="/workspace/third_party/peract2_bimanual_probe_20260324")
     parser.add_argument("--print-train-command", action="store_true")
     args = parser.parse_args()
 

code/reveal_vla_bimanual/sim_rlbench/task_resolver.py

@@ -0,0 +1,11 @@
+from __future__ import annotations
+
+from rlbench.backend.utils import task_file_to_task_class
+
+
+def is_bimanual_task_name(task_name: str) -> bool:
+    return task_name.startswith("bimanual_")
+
+
+def resolve_task_class(task_name: str):
+    return task_file_to_task_class(task_name, bimanual=is_bimanual_task_name(task_name))

environment/run_peract2_13_rollouts.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="${ROOT_DIR:-/workspace}"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+BUNDLE_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+PROJECT_DIR="${PROJECT_DIR:-${ROOT_DIR}/reveal_vla_bimanual}"
+MAMBA_BIN="${ROOT_DIR}/.tools/micromamba/bin/micromamba"
+MAMBA_ROOT_PREFIX="${ROOT_DIR}/.micromamba"
+ENV_PREFIX="${ROOT_DIR}/envs/rlbench"
+DISPLAY_NUM="${DISPLAY_NUM:-99}"
+DISPLAY=":${DISPLAY_NUM}"
+COPPELIA_DIR="${COPPELIASIM_ROOT:-${ROOT_DIR}/assets/coppeliasim_v4_1_0}"
+RUNTIME_DIR="${ROOT_DIR}/runtime"
+CHECKPOINT="${CHECKPOINT:-${BUNDLE_ROOT}/artifacts/outputs/interaction_debug/proxy_interaction_state_clip_actionhist/checkpoint_best.pt}"
+OUTPUT_ROOT="${OUTPUT_ROOT:-${BUNDLE_ROOT}/artifacts/outputs/interaction_debug}"
+RUN_NAME_PREFIX="${RUN_NAME_PREFIX:-peract2_13_rollout}"
+
+mkdir -p "${RUNTIME_DIR}"
+chmod 700 "${RUNTIME_DIR}"
+
+ensure_rlbench_display() {
+  if DISPLAY="${DISPLAY}" xdpyinfo >/dev/null 2>&1; then
+    return 0
+  fi
+  ROOT_DIR="${ROOT_DIR}" DISPLAY_NUM="${DISPLAY_NUM}" "${PROJECT_DIR}/scripts/start_rlbench_x.sh"
+}
+
+run_in_rlbench_env() {
+  local driver_version=""
+  local driver_branch=""
+  local shim_ld=""
+  if command -v nvidia-smi >/dev/null 2>&1; then
+    driver_version="$(nvidia-smi --query-gpu=driver_version --format=csv,noheader | head -n1 || true)"
+    driver_branch="${driver_version%%.*}"
+    if [[ -n "${driver_branch}" && -d "${ROOT_DIR}/system_shims/nvidia${driver_branch}/usr/lib/x86_64-linux-gnu" ]]; then
+      shim_ld="${ROOT_DIR}/system_shims/nvidia${driver_branch}/usr/lib/x86_64-linux-gnu"
+    fi
+  fi
+  env \
+    DISPLAY="${DISPLAY}" \
+    COPPELIASIM_ROOT="${COPPELIA_DIR}" \
+    XDG_RUNTIME_DIR="${RUNTIME_DIR}" \
+    LD_LIBRARY_PATH="${COPPELIA_DIR}${shim_ld:+:${shim_ld}}${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}" \
+    QT_QPA_PLATFORM_PLUGIN_PATH="${COPPELIA_DIR}" \
+    "${MAMBA_BIN}" run -r "${MAMBA_ROOT_PREFIX}" -p "${ENV_PREFIX}" "$@"
+}
+
+ensure_rlbench_display
+
+run_in_rlbench_env python -m eval.run_peract2_task_sweep \
+  --checkpoint "${CHECKPOINT}" \
+  --output-root "${OUTPUT_ROOT}" \
+  --run-name-prefix "${RUN_NAME_PREFIX}" \
+  --headless \
+  --chunk-commit-steps 4 \
+  "$@"

environment/setup_same_machine.sh

@@ -8,6 +8,7 @@ BUNDLED_PROJECT_DIR="${BUNDLE_ROOT}/code/reveal_vla_bimanual"
 PROJECT_DIR="${PROJECT_DIR:-${ROOT_DIR}/reveal_vla_bimanual}"
 THIRD_PARTY_DIR="${ROOT_DIR}/third_party"
 INSTALL_ENV_B="${INSTALL_ENV_B:-1}"
+INSTALL_LEGACY_LAYOUT="${INSTALL_LEGACY_LAYOUT:-0}"
 
 echo "Using ROOT_DIR=${ROOT_DIR}"
 echo "Using BUNDLE_ROOT=${BUNDLE_ROOT}"
@@ -39,10 +40,16 @@ mkdir -p "${THIRD_PARTY_DIR}"
 ensure_project_dir
 
 echo "1/6 Sync pinned third-party repos"
-sync_repo "${THIRD_PARTY_DIR}/peract_bimanual" "https://github.com/markusgrotz/peract_bimanual.git" "bb0232a6ba3fe116566e9568f0c7af980ed6703d"
-sync_repo "${THIRD_PARTY_DIR}/RLBench" "https://github.com/markusgrotz/RLBench.git" "8af748c51287989294e00c9c670e3330a0e35ed5"
-sync_repo "${THIRD_PARTY_DIR}/PyRep" "https://github.com/markusgrotz/PyRep.git" "b8bd1d7a3182adcd570d001649c0849047ebf197"
-sync_repo "${THIRD_PARTY_DIR}/YARR" "https://github.com/markusgrotz/YARR.git" "6822ff78602c77878b27d4cfe759ce029c67bffb"
+sync_repo "${THIRD_PARTY_DIR}/peract2_bimanual_probe_20260324" "https://github.com/markusgrotz/peract_bimanual.git" "bb0232a6ba3fe116566e9568f0c7af980ed6703d"
+sync_repo "${THIRD_PARTY_DIR}/RLBench_peract2_20260324" "https://github.com/markusgrotz/RLBench.git" "8af748c51287989294e00c9c670e3330a0e35ed5"
+sync_repo "${THIRD_PARTY_DIR}/PyRep_peract2_20260324" "https://github.com/markusgrotz/PyRep.git" "b8bd1d7a3182adcd570d001649c0849047ebf197"
+sync_repo "${THIRD_PARTY_DIR}/YARR_peract2_20260324" "https://github.com/markusgrotz/YARR.git" "6822ff78602c77878b27d4cfe759ce029c67bffb"
+if [[ "${INSTALL_LEGACY_LAYOUT}" == "1" ]]; then
+  sync_repo "${THIRD_PARTY_DIR}/peract_bimanual" "https://github.com/markusgrotz/peract_bimanual.git" "1cc074b1730f52a7fd35d069d8c9099d18718bca"
+  sync_repo "${THIRD_PARTY_DIR}/RLBench" "https://github.com/markusgrotz/RLBench.git" "02720bba4c73fe02eb75df946b8791b806028a9d"
+  sync_repo "${THIRD_PARTY_DIR}/PyRep" "https://github.com/markusgrotz/PyRep.git" "b8bd1d7a3182adcd570d001649c0849047ebf197"
+  sync_repo "${THIRD_PARTY_DIR}/YARR" "https://github.com/markusgrotz/YARR.git" "6822ff78602c77878b27d4cfe759ce029c67bffb"
+fi
 
 echo "2/6 Install or update Env A (RLBench / PerAct2)"
 ROOT_DIR="${ROOT_DIR}" "${PROJECT_DIR}/scripts/setup_env_a_rlbench.sh"
@@ -67,6 +74,7 @@ echo "  INSTALL_ISAACSIM=1 ${PROJECT_DIR}/scripts/setup_env_b_reveal.sh"
 echo
 echo "Recommended validation commands:"
 echo "  DISPLAY=:99 glxinfo -B"
-echo "  /workspace/.tools/micromamba/bin/micromamba run -r /workspace/.micromamba -p /workspace/envs/rlbench python -m sim_rlbench.launch_smoke --headless"
+echo "  /workspace/.tools/micromamba/bin/micromamba run -r /workspace/.micromamba -p /workspace/envs/rlbench python -m sim_rlbench.launch_smoke --task bimanual_push_box --headless"
+echo "  ${BUNDLE_ROOT}/environment/run_peract2_13_rollouts.sh --episodes-per-task 1 --episode-length 20"
 echo "  /workspace/.tools/micromamba/bin/micromamba run -r /workspace/.micromamba -p /workspace/envs/rlbench python -m sim_reveal.isaac_smoke"
 echo "  ${BUNDLE_ROOT}/environment/validate_same_machine.sh"

environment/upstream_revisions.txt

@@ -1,11 +1,27 @@
-/workspace/third_party/peract_bimanual
+/workspace/third_party/peract2_bimanual_probe_20260324
 repo_url=https://github.com/markusgrotz/peract_bimanual.git
 commit=bb0232a6ba3fe116566e9568f0c7af980ed6703d
 
-/workspace/third_party/RLBench
+/workspace/third_party/RLBench_peract2_20260324
 repo_url=https://github.com/markusgrotz/RLBench.git
 commit=8af748c51287989294e00c9c670e3330a0e35ed5
 
+/workspace/third_party/PyRep_peract2_20260324
+repo_url=https://github.com/markusgrotz/PyRep.git
+commit=b8bd1d7a3182adcd570d001649c0849047ebf197
+
+/workspace/third_party/YARR_peract2_20260324
+repo_url=https://github.com/markusgrotz/YARR.git
+commit=6822ff78602c77878b27d4cfe759ce029c67bffb
+
+/workspace/third_party/peract_bimanual
+repo_url=https://github.com/markusgrotz/peract_bimanual.git
+commit=1cc074b1730f52a7fd35d069d8c9099d18718bca
+
+/workspace/third_party/RLBench
+repo_url=https://github.com/markusgrotz/RLBench.git
+commit=02720bba4c73fe02eb75df946b8791b806028a9d
+
 /workspace/third_party/PyRep
 repo_url=https://github.com/markusgrotz/PyRep.git
 commit=b8bd1d7a3182adcd570d001649c0849047ebf197

environment/validate_same_machine.sh

@@ -62,8 +62,8 @@ DISPLAY="${DISPLAY}" glxinfo -B
 
 echo
 echo "RLBench import smoke"
-run_in_rlbench_env python -m sim_rlbench.smoke_test --print-train-command
+run_in_rlbench_env python -m sim_rlbench.smoke_test --print-train-command --upstream-root "${ROOT_DIR}/third_party/peract2_bimanual_probe_20260324"
 
 echo
-echo "RLBench launch smoke"
-run_in_rlbench_env python -m sim_rlbench.launch_smoke --task open_drawer --resolution 128 --headless
+echo "RLBench launch smoke: bimanual_push_box"
+run_in_rlbench_env python -m sim_rlbench.launch_smoke --task bimanual_push_box --resolution 128 --headless