Superseded simulation-only staging PR

simulation/rmbench/molmoact2_eval_assets/manifest.json

@@ -10,6 +10,7 @@
     "revision": "1cfa9a7208912126459214e8b04321603b3df60c"
   },
   "schema_version": 1,
+  "source_run": "mit0jkx7",
   "stage2_metadata": [
     "stage2/meta/episodes.jsonl",
     "stage2/meta/memory_layout.json"

simulation/robocasa/600/README.md

@@ -29,8 +29,7 @@ python scripts/reproduce/extract_dataset_store.py \
 ```
 
 See the manifests for per-shard SHA256 checksums. The exact patched simulator
-source is distributed in `Amazon-FAR/imprint_datasets`, path
-`simulation/robocasa/environment`.
+source is distributed separately as `yilin-wu/imprint-robocasa-environment`.
 
 The corresponding code, launch configurations, and validation tools are
 published in the IMPRINT GitHub repository.

simulation/robocasa/environment/README.md

@@ -16,18 +16,17 @@ Included:
 
 - 18 migrated memory-task classes
 - legacy RoboCasa 0.2 scene-index compatibility
-- bundled legacy scene metadata required by the memory tasks
 - recorded XML/state replay fixes
 - task registration and terminal success logic
 - pinned Python and dependency manifests
+- environment tests and migration notes
 
 Archive:
 
 ```text
 robocasa-memory-setup.tar.gz
-SHA256 12cb7a92087298e6e9f299070d294655e54f1340ef578ca5e0581c885a79d9fc
+SHA256 ea113f5e102a59d8c7b893dbc9f4dfe3e993cc5dd1a0ec5f4173d9012061bd91
 ```
 
-The large RoboCasa kitchen object and texture assets are not embedded. Use the
-IMPRINT environment setup helper after extraction; it downloads those assets
-from their public upstream sources.
+The large RoboCasa kitchen and legacy replay assets are not embedded. Follow
+`SETUP_AND_FIX_NOTES.md` after extraction.

simulation/robocasa/environment/SETUP_AND_FIX_NOTES.md

@@ -0,0 +1,159 @@
+# RoboCasa × ReMemBench Memory Tasks — Full Setup & Fix Notes
+
+This package is the complete, self-contained setup for the **migrated ReMemBench
+memory tasks running inside official RoboCasa 1.0.1 (robocasa-new / robocasa365)**,
+plus the datasets pipeline (replay → regenerate → LeRobot convert → merge with
+human300) and the Robo-Dopamine reward-annotation pipeline built on top.
+
+It captures **what the mixture between RoboCasa and ReMemBench was, why it broke, and
+exactly how it was fixed**, together with all the code and the reproducible
+environment manifest.
+
+---
+
+## 0. TL;DR — the environment
+
+Everything below was built and run in ONE uv-managed environment:
+
+- **`~/robocasa-new/.venv-mem`** — CPython **3.10.20**
+- **robocasa 1.0.1** (editable install = the code in this package)
+- **robosuite 1.5.2**, **mujoco 3.3.1** (CPU physics, EGL rendering)
+- torch 2.7.1, numpy 2.2.5, h5py 3.16.0, imageio 2.37.3, **lerobot 0.3.3**
+- Host GPU: NVIDIA RTX PRO 5000 (Blackwell, sm_120) — rendering uses `MUJOCO_GL=egl`.
+
+Exact pins: `requirements-venv-mem.freeze.txt` (full `pip freeze`) + `PYTHON_VERSION.txt`.
+
+Recreate:
+```bash
+uv venv --python 3.10 .venv-mem
+uv pip install -r requirements-venv-mem.freeze.txt --python .venv-mem/bin/python
+uv pip install -e . --python .venv-mem/bin/python     # this repo = robocasa 1.0.1
+```
+
+The Robo-Dopamine reward VLM runs in a SEPARATE env (`~/robometer/.venv-robodopamine`,
+vLLM) — see §5 and the robometer package notes.
+
+---
+
+## 1. What the "mixture" was
+
+The memory tasks originated in **ReMemBench**, a fork of **robocasa 0.2.0** (used in the
+HALO project). The goal was to run those tasks in **official robocasa 1.0.1**
+(a.k.a. robocasa-new / robocasa365) so a single pi0.5/openpi policy can train and eval
+on memory + non-memory tasks in one ecosystem.
+
+Two distinct things had to be reconciled:
+
+1. **Environment classes** — 18 `Mem*` task definitions written against the 0.2.0 API.
+2. **Recorded demonstrations** — HDF5 demos recorded under 0.2.0 / robosuite 1.x, which
+   must replay faithfully on the 1.0.1 envs.
+
+---
+
+## 2. How it was fixed — three layers
+
+Full detail is in the three notes docs shipped in this package under
+`docs/superpowers/`. Summary of each fix layer:
+
+### Layer 1 — Env-class port  (`MEMORY_TASKS_MIGRATION_NOTES.md`)
+- Ported 18 `Mem*` classes to `robocasa/environments/kitchen/memory/memory_env.py`.
+- Ported ReMemBench object groups + cooking-timing tables into
+  `robocasa/models/objects/kitchen_objects.py`, plus 2 `object_utils.py` helpers.
+- Registered the tasks so 1.0.1's env registry can instantiate them.
+
+### Layer 2 — Legacy-scene replay  (`LEGACY_REPLAY_NOTES.md` — supersedes L1's "blocked" conclusion)
+Three sequential blockers between 0.2.0 demos and 1.0.1 envs:
+- **Scene index conflict:** 0.2.0 used 0-indexed named style/layout enums; 1.0.1
+  renumbered to 1-indexed numeric (`STYLE001=1…`). Recorded `style_id=0` → `KeyError`.
+  Fixed by remapping the migrated 0.2.0 scene indices to registered legacy entries
+  (`scene_registry` + `ep_meta` remap on every `reset_to`).
+- **`ep_meta` object-config schema drift:** 0.2.0 keys (`obj_registries`,
+  `obj_instance_split`, nested `object_scale`) rejected by the 1.0.1 cfg validator.
+  Fixed by tolerating/normalizing legacy keys during replay.
+- **Fixture asset gap:** legacy 0.2.0 fixture assets missing in 1.0.1. Fixed by a
+  reproducible fetch of the legacy assets (`scripts/replay/fetch_legacy_assets.py`).
+- **Render fix:** hide baked-in `success_region` / debug-visualization goal sites so
+  they don't appear in rendered frames (`_hide_debug_sites`).
+
+### Layer 3 — Dataset regeneration  (`REGENERATED_DATASET_NOTES.md`)
+- **State-faithful replay:** restore each recorded MuJoCo state via `reset_to`,
+  re-derive obs + render cameras, re-write in the original two-file ReMemBench format
+  with migrated version stamps. Not a fresh re-collection — recorded actions preserved.
+- Persisted container-lift correction; all demos end `is_success=True`.
+
+---
+
+## 3. Re-render for the unified dataset (256×256, 3 cameras)
+
+After the mixture was resolved, the memory demos were re-rendered to match human300 so
+they could be merged into one LeRobot dataset. Design + plan:
+`docs/superpowers/specs/2026-07-06-unified-memory-human300-dataset-design.md`.
+
+- **Re-render:** `scripts/regen_dataset/` (`regenerate_all.py` → `regenerate_demo.py`
+  → `write_hdf5.py`). Sets `CAMERAS = [robot0_agentview_left, robot0_agentview_right,
+  robot0_eye_in_hand]` (center dropped) and `camera_heights=camera_widths=256`.
+  EGL default, ~300 fps, resumable skip-if-exists.
+  ```bash
+  MUJOCO_GL=egl .venv-mem/bin/python -m scripts.regen_dataset.regenerate_all \
+    --out-dir regenerated_dataset_256   # 256px / 3-cam
+  ```
+- **Memory → LeRobot v2.1:** `scripts/merge_dataset/convert_memory_to_lerobot.py` —
+  real 3-cam mapping (`agentview_left`→left, `agentview_right`→right, `eye_in_hand`→
+  wrist), **vertical-flip** (memory obs are bottom-up), h264 video to match human300,
+  `state[16]`/`action[12]`, `PandaOmron_modality.json`. Output: `memory_lerobot/<task>/lerobot`.
+- **human300 download:** `scripts/merge_dataset/download_human300.sh`.
+- **Physical merge:** `scripts/merge_dataset/merge_lerobot_datasets.py` — concatenates
+  313 task dirs (13 memory + 300 human300), contiguous re-index of
+  `episode_index`/`index`/`task_index` (+ `annotation.human.*`), dedup tasks.jsonl,
+  frame-weighted stats aggregation.
+
+**Camera facts (verified, model consumes 3 cams):** `base_0_rgb`←agentview_left,
+`right_wrist_0_rgb`←agentview_right, `left_wrist_0_rgb`←eye_in_hand. Stored 256×256;
+model resizes to 224×224 at load. agentview_center is NOT used.
+
+The public IMPRINT training store is under
+`simulation/robocasa/600` in `Amazon-FAR/imprint_datasets`. This customized
+environment package is under `simulation/robocasa/environment` in the same
+repository.
+
+---
+
+## 4. Package layout
+
+```
+SETUP_AND_FIX_NOTES.md              # this file
+PYTHON_VERSION.txt                  # 3.10.20
+requirements-venv-mem.freeze.txt    # full pip freeze of .venv-mem
+robocasa/                           # robocasa 1.0.1 source + memory/ env package
+  environments/kitchen/memory/      # THE 18 Mem* task classes (memory_env.py)
+  models/objects/kitchen_objects.py # ported object groups + cooking timings
+scripts/
+  regen_dataset/                    # re-render (replay -> hdf5)
+  merge_dataset/                    # convert to LeRobot + download human300 + merge
+  replay/                           # legacy replay, key-moment detection, wash split
+tests/memory/, tests/merge/         # registration/smoke/replay/convert verification
+docs/superpowers/
+  MEMORY_TASKS_MIGRATION_NOTES.md   # Layer 1
+  LEGACY_REPLAY_NOTES.md            # Layer 2
+  REGENERATED_DATASET_NOTES.md      # Layer 3
+  specs/ plans/                     # design + implementation plans (dated)
+```
+
+---
+
+## 5. Reward annotations (Robo-Dopamine, separate env)
+
+The 4-segment wash-and-return reward pipeline lives in the separate
+`robometer` project and is not included in this environment archive. The
+training store published with IMPRINT already contains the reward annotations
+used by the released configuration.
+
+---
+
+## 6. Git provenance
+
+The portable evaluator, setup script, and environment manifest are maintained
+in `Amazon-FAR/imprint`. The archived commit history traces the exact fix
+order: object groups → env port → legacy scene remap → ep_meta tolerance →
+asset fetch → debug-site hiding → replay → regenerate → 256/3-cam
+parameterization → LeRobot conversion → human300 merge.

simulation/robocasa/environment/robocasa-memory-setup.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12cb7a92087298e6e9f299070d294655e54f1340ef578ca5e0581c885a79d9fc
-size 2181680
+oid sha256:ab65ddb296584f81261c1d95a154e00ae73b3d874c8d1306b9c029a38b734eb5
+size 41732524