Document Revo3 reference, policy I/O and deployment integration

README.md

@@ -0,0 +1,17 @@
+# DexSys inference bundles
+
+Exported actors, matching references and simulation scenes for
+[DexSys Runtime](https://github.com/Chuan-10/dexsys-runtime). Runtime code and
+installation instructions are on GitHub; weights and assets are hosted here.
+
+| Bundle | Scope |
+| --- | --- |
+| [Flexiv + right Revo3](revo3/flexiv-right-revo3-125315-w4-seed42-u100/README.md) | 125315 window 4; `pose` and `pose_contact`, both seed42 / update100; includes the required input reference |
+| [Flexiv + right Sharpa](https://huggingface.co/datasets/Chrisss/DexSys/tree/main/baseline/flexiv-right-sharpa-125315-seed43-u1500) | 125315 single-reference baseline, seed43 / update1500; [runtime instructions](https://github.com/Chuan-10/dexsys-runtime#sharpa-baseline-assets) |
+
+Each bundle contains its actor, input/output contract, reference, scene and
+checksums. Follow the linked instructions for the selected bundle;
+hand types, joint orders and observation layouts are not interchangeable.
+These packages support inference and simulation verification. Live perception,
+robot SDK integration and hardware calibration are not included. Preserve the
+asset notices within each bundle.

revo3/flexiv-right-revo3-125315-w4-seed42-u100/README.md

@@ -1,46 +1,85 @@
 # Flexiv + right Revo3: 125315 window 4
 
-Two deterministic **seed42 / update100** policies: `policies/pose` and
-`policies/pose_contact`. Use with [DexSys Runtime v0.3.0](https://github.com/Chuan-10/dexsys-runtime/tree/v0.3.0).
-The GitHub repository contains inference and simulation code only.
+Two deterministic **seed42 / update100** actors, `pose` and `pose_contact`, share
+one fixed fourth-window reference. Install
+[DexSys Runtime](https://github.com/Chuan-10/dexsys-runtime#install) (code v0.3.0).
+This is an inference/simulation bundle; hardware integration is required separately.
 
-| Content | Purpose |
+## Included files
+
+| Path | Content |
 | --- | --- |
-| `policies/*/actor.npz`, `contract.json` | 522 → 128 ELU → 128 ELU → 28 actor, frozen normalization and explicit joint/feature order |
-| `case/scene` | Exact collision/visual meshes, MJCF and numeric robot profile |
-| `case/reference.npz`, `task.json`, `candidates.json` | Retargeted joint/object reference, one motion window and derived contact anchors |
-| `case/initialization/state.npz` | Executed prefix qpos, qvel and last sent command; mandatory cold-reset state |
-| `case/window.json`, `case_manifest.json` | Source timing and current/original SHA-256 links |
-| `replays/two-policies-fixed-camera.mp4` | Audited batch00 / world0 for each policy, fixed workspace and overview cameras |
-| `VALIDATION.json`, `SHA256SUMS` | Numeric implementation and execution checks, file integrity |
+| `policies/{pose,pose_contact}/actor.npz`, `contract.json` | Actor weights, frozen normalization, exact input/joint order and limits |
+| `case/reference.npz` | Required input reference: 242 frames / 8.0333 s, timestamps, `q_target [242,28]`, joint names, four object IDs, `reference_objects_world [242,4,4,4]` and availability masks |
+| `case/scene` | Robot/table/object MJCF, collision/visual meshes and numeric robot profile |
+| `case/{task.json,candidates.json,window.json}` | Task window, finger/object-local anchor correspondences and original source timing |
+| `case/initialization/state.npz` | Executed-prefix `qpos`, `qvel` and previous command `ctrl` for simulation reset |
+| `case/case_manifest.json`, `SHA256SUMS`, `VALIDATION.json` | File integrity and release validation |
+| `replays/two-policies-fixed-camera.mp4` | Audited fixed-camera comparison; numeric command traces are generated locally |
+
+`q_target` is the **nominal IK reference**, not the PPO-corrected command. No raw
+MANO data or additional IK solve is needed for inference. The local reference
+clock is 0–8.0333 s; the active window is 0.85–5.9667 s, with original source
+offset +29.8 s. The runtime interpolates the 30 Hz reference to 100 Hz.
+
+## Inputs and outputs
+
+| Interface | Contract |
+| --- | --- |
+| Input | `float32 [batch,522]`: live robot state, current/future reference and FK goals, actual/goal object states, contacts/masks, candidate distances, previous command and task/contact history |
+| Fixed scene slots | `G21_1`–`G21_4`, all retained; window 4 manipulates `G21_1` |
+| Coordinates | Scene/world positions in metres, joint angles in radians, quaternions `wxyz`; object velocities use MuJoCo free-joint semantics |
+| Actor | 522 → 128 ELU → 128 ELU → 28; packaged normalization applied by `predict` |
+| Raw output | 28 dimensionless residual means, without sampling or tanh |
+| Final output | 7 Flexiv + 21 Revo3 joint position targets: next nominal + scaled/clipped residual, then position/rate limits |
+| Residual/control | Arm ±0.08 rad, hand ±0.12 rad; 100 Hz commands; simulation uses PD with arm gravity compensation |
+
+Use the exact field/joint order in `contract.json`. The policy needs live object
+and contact estimates for hardware use; stored goals cannot replace measured
+state. Keep all object slots. Unknown contact is not confirmed no-contact.
+Preserve continuous hold/history updates, including all four `task_stop_history`
+channels for both actors. See the
+[runtime input contract](https://github.com/Chuan-10/dexsys-runtime#policy-inputs-and-outputs).
+
+## Run and integrate
 
-Run from the downloaded bundle directory after installing the runtime:
+After installation and download, run from this bundle directory:
 
 ```bash
 dexsys-runtime verify-case --case case
 dexsys-runtime inspect --policy policies/pose_contact
 dexsys-runtime simulate --case case --policy policies/pose_contact --output outputs/cpu
-# Optional pinned native mjlab environment:
+# Use policies/pose for the other actor, or --zero-residual for its control.
+# Optional pinned mjlab environment:
 dexsys-runtime simulate --case case --policy policies/pose --backend mjlab --device cuda:0 --output outputs/gpu
 ```
 
-Add `--zero-residual` for the matching control; use a new output directory for each run.
-Prepared reset and four stop-history channels are required. The 21-joint Revo3
-hand is not compatible with the 22-joint Sharpa action order. Residual budgets,
-force limits, arm–table collision and final task thresholds are unchanged.
-
-Both exported actors completed the task in independent CPU and native GPU runs.
-The two zero-residual GPU controls stopped at 2.40 s with task failure. Independent
-CPU geometry audits passed all 2,092 saved GPU states. Actor outputs were exact
-on 2,092 shared-state observations; constrained commands were exact on 2,088
-transitions. These checks include the failed controls.
-
-Archived source evaluations achieved 64/64 task passes for each actor from one
-fixed initial scene. This measures repeatability, not generalization. These are
-partial-training models for the isolated fourth window; the original servo
-tracking gate still fails. No hardware deployment, arbitrary-object perception,
-or full four-window task success is claimed.
-
-The bundle contains no critic, optimizer, training launcher, raw human motion,
-research notes, private paths or credentials. Preserve `NOTICE.md` and upstream
-asset conditions.
+Use a new output directory for each run. Inspect `summary.json`: physical
+validity, task success and servo tracking are separate results.
+
+| Deployment mode | What to use / what is still required |
+| --- | --- |
+| Open-loop joint replay | `simulate` saves `trace.npz`: `sent[:,0,:]` is the constrained command; `reference_time[:,0]` is its target-reference timestamp. Bind names from `contract.json` and implement SDK mapping/timing. Playing `q_target` alone omits PPO corrections |
+| Closed-loop policy | At 100 Hz, build the complete live observation, call `predict`, then the persistent `JointTargetController`; send `command.sent` through a hardware adapter and update task/contact history |
+
+The package has no robot driver, live perception/contact adapter or hardware
+calibration. Deployment must align joint order/zero/sign/units, scene coordinates,
+servo behavior, initial-pose transition and reference/history reset. Simulator
+`qpos/qvel` snapshots are not hardware reset commands. Check the actual start
+transition and commands against hardware limits and singularities. This release
+uses the original arm reference, without the experimental null-space refinement.
+See [hardware integration](https://github.com/Chuan-10/dexsys-runtime#hardware-integration).
+
+## Validation scope
+
+Both actors completed the task in independent CPU and native GPU runs. The two
+zero-residual GPU controls stopped at 2.40 s with task failure. CPU geometry audits
+passed all 2,092 saved GPU states; actor outputs were exact on 2,092 observations
+and constrained commands on 2,088 transitions, including failed controls.
+
+Archived 64/64 task passes per actor repeat one fixed scene and initialization;
+they do not establish generalization. These are partial-training models for
+window 4; the original servo-tracking gate still fails. Hardware deployment and
+full four-window success have not been established. No critic, optimizer, training
+launcher, raw human motion or private research notes are included. Preserve
+`NOTICE.md` and upstream asset conditions.

revo3/flexiv-right-revo3-125315-w4-seed42-u100/SHA256SUMS

@@ -1,5 +1,5 @@
 34c3ad3dc3988c29e95062dd6577e80689721e045f056b01d6b7e62fcc06ee8b  NOTICE.md
-d7078baf612b7f75c0f301c3e4bb429d8468967f492b61a09fe55c9719d900aa  README.md
+e86278c33da16893c43c448a36261e0b40d29261a1f0e924c5e365f30bbfede1  README.md
 83e23b288033486a6b2637e8481552e145e3b48c7865194e76965f646edb2e99  VALIDATION.json
 e498ce3f98d3944f553146ec6e3ece4d8ac741e3e342daa2d6ddae9bb49f144a  case/candidates.json
 218077c482465df6cd4fbde78c847c106705ca1c659342d7b43faea3ca7a681d  case/case_manifest.json