Add Astra trajectory augmentation guide

README.md

@@ -3,6 +3,7 @@ pretty_name: Sharpa Artifacts
 tags:
 - robotics
 - dexterous-manipulation
+- reinforcement-learning
 - retargeting
 - hdf5
 - mano
@@ -10,47 +11,92 @@ tags:
 
 # Sharpa Artifacts
 
-Public, versioned machine artifacts for the Sharpa functional-reorientation pipeline.
-Source code and download validators live in
+Public, versioned machine artifacts for the Sharpa functional-reorientation
+pipeline. Source code and download validators live in
 [`ai4ce/dex_retarget`](https://github.com/ai4ce/dex_retarget).
 
+## Start here: Astra trajectory augmentation
+
+The repository now includes eight Sharpa Stage3 AutoDR residual-teacher
+checkpoint **candidates** for generating physics-executed recovery trajectories.
+Give an Astra agent the following guide as its execution handoff:
+
+- [Astra guide: use the Phone8 teachers for trajectory augmentation](https://huggingface.co/datasets/FlappyHimself/sharpa-artifacts/blob/main/guides/phone8-stage3-autodr-candidates-v1/ASTRA_GUIDE.md)
+- [Immutable Phone8 checkpoint release](https://huggingface.co/datasets/FlappyHimself/sharpa-artifacts/tree/942f7cd025ae050823a817c9aff6cb7cb810d009/objects/phone/releases/phone8-stage3-autodr-candidates-v1)
+
+The intended loop is:
+
+```text
+matching Stage2 reference
+  -> bounded hand/object tweak
+  -> residual-teacher recovery in Isaac/PhysX
+  -> functional-pose evaluation
+  -> immutable rollout + action chain + diagnostics + 2x3 video
+  -> optional accepted child tweak
+```
+
+This release contains the teachers and collection contract. It does **not**
+claim that the proposed augmented rollout dataset has already been collected.
+
 ## Layout
 
 Artifacts are organized by object and immutable release:
 
 ```text
 objects/<object>/releases/<release-id>/
+guides/<release-id>/
 ```
 
-The current Stage3-compatible release is
-`objects/phone/releases/phone20-stage12-v2`. It contains 20 accepted Stage1
-NPZ files, 20 final Sharpa `retargeted_moving28_60fps.h5` trajectories, and the
-human/canonical H5, QA video, manifests, GoalContracts, and handoffs required by
-the current strict Stage3 provenance contract.
+## Releases
 
-`release_manifest.json` and `checksums.sha256` define every released byte.
-Consumers should pin an exact repository revision and verify the release
-manifest before use.
+### `phone8-stage3-autodr-candidates-v1`
 
-## Releases
+- Eight per-phone Stage3 residual-teacher checkpoint candidates: `phone_02`,
+  `phone_03`, `phone_05`, `phone_07`, `phone_10`, `phone_14`, `phone_17`, and
+  `phone_18`.
+- Each episode includes its checkpoint, Training Card, checkpoint receipt and
+  evaluation summary.
+- The release also includes a Data Card, collection specification, manifest and
+  SHA256 inventory.
+- Status: `CHECKPOINT_RELEASE_READY_CANDIDATES_NOT_PROMOTED`.
+
+### `phone20-stage12-v2`
+
+- Current Stage3-compatible Stage1/Stage2 release.
+- Contains 20 accepted Stage1 NPZ files, 20 final Sharpa
+  `retargeted_moving28_60fps.h5` trajectories, human/canonical H5 files,
+  GoalContracts, QA videos, manifests and handoffs.
+
+### `phone20-stage12-v1`
 
-- `phone20-stage12-v2`: current Stage3-compatible layout; preserves each
-  handoff's original `inputs/`, `stage2/`, `contracts/`, and `qa/` paths.
-- `phone20-stage12-v1`: historical first upload; usable for direct artifact
-  inspection but superseded for Stage3 because it renamed `inputs/` to `stage1/`
-  and omitted the handoff-bound QA video.
+- Historical first upload.
+- Superseded for Stage3 because it renamed `inputs/` to `stage1/` and omitted
+  the handoff-bound QA video.
+
+## Consumer contract
+
+Pin an exact Hugging Face revision and verify `release_manifest.json` plus
+`checksums.sha256` before use. A checkpoint must be paired with the matching
+episode's Stage2 H5, GoalContract, Training Card and runtime identity; similar
+filenames are not interchangeable provenance.
+
+Checkpoint availability is not policy promotion. These candidates have not
+completed the full P2 robustness bank, automatic drop detection and human video
+review required for a robust-teacher claim. See the per-episode evaluation
+summary before selecting a source.
 
 ## Data boundary
 
-This repository contains derived research artifacts. It does not contain raw
-Vicon/MANUS capture, checkpoints, teacher rollouts, behavior-cloning datasets,
-or training logs. Public access does not grant rights to unrelated third-party
-assets; users remain responsible for their applicable licenses.
+This repository contains derived research artifacts and selected checkpoint
+candidates. It does not contain raw Vicon/MANUS capture, the proposed new
+augmentation rollout bytes, full training logs, a self-contained Isaac image,
+or unrelated third-party assets. Public access does not grant rights to
+third-party assets; consumers remain responsible for applicable licenses.
+
+## Pinned identities
 
-## Current Stage3-compatible identity
+| Release | Artifact revision | Manifest SHA256 | Status |
+|---|---|---|---|
+| `phone8-stage3-autodr-candidates-v1` | `942f7cd025ae050823a817c9aff6cb7cb810d009` | `d8b0c52ebb5155eef17b033167b600114efd7bca22278e966e0136c80036b277` | checkpoint candidates published; rollouts not materialized |
+| `phone20-stage12-v2` | `7c3a60460e2c953c929fbfffdc1ab5a9a8607851` | `84bd39b42ba3cdd0f9985fb4d6e1c002654d051dd7666b9fe9350b049b9b4149` | Stage3-compatible Stage1/Stage2 release |
 
-- Release: `phone20-stage12-v2`
-- Artifact commit: `7c3a60460e2c953c929fbfffdc1ab5a9a8607851`
-- Release manifest SHA256: `84bd39b42ba3cdd0f9985fb4d6e1c002654d051dd7666b9fe9350b049b9b4149`
-- Source producer commit: `22d7dd4ae9dd782625464788e5741a0565311c77`
-- Source state: clean; all 20 handoffs record `STAGE2_HANDOFF_ACCEPTED`

catalog.json

@@ -1,7 +1,20 @@
 {
   "canonical_release": "phone20-stage12-v2",
+  "featured_teacher_release": "phone8-stage3-autodr-candidates-v1",
   "public": true,
   "releases": [
+    {
+      "artifact_revision": "942f7cd025ae050823a817c9aff6cb7cb810d009",
+      "checkpoint_count": 8,
+      "episode_count": 8,
+      "guide_path": "guides/phone8-stage3-autodr-candidates-v1/ASTRA_GUIDE.md",
+      "manifest_sha256": "d8b0c52ebb5155eef17b033167b600114efd7bca22278e966e0136c80036b277",
+      "member_count": 37,
+      "object": "phone",
+      "path": "objects/phone/releases/phone8-stage3-autodr-candidates-v1",
+      "release_id": "phone8-stage3-autodr-candidates-v1",
+      "status": "CHECKPOINT_RELEASE_READY_CANDIDATES_NOT_PROMOTED"
+    },
     {
       "artifact_revision": "7c3a60460e2c953c929fbfffdc1ab5a9a8607851",
       "episode_count": 20,

guides/phone8-stage3-autodr-candidates-v1/ASTRA_GUIDE.md

@@ -0,0 +1,251 @@
+# Astra guide: Sharpa Phone8 teacher trajectory augmentation
+
+This guide is an execution handoff for an Astra coding/research agent. It
+describes how to use the published Sharpa residual-teacher checkpoint
+**candidates** to prepare and collect diverse physics-executed recovery
+trajectories.
+
+## Fixed release identity
+
+| Item | Identity |
+|---|---|
+| Hugging Face repository | `FlappyHimself/sharpa-artifacts` (`dataset`, public) |
+| Teacher release | `phone8-stage3-autodr-candidates-v1` |
+| Teacher artifact revision | `942f7cd025ae050823a817c9aff6cb7cb810d009` |
+| Release manifest SHA256 | `d8b0c52ebb5155eef17b033167b600114efd7bca22278e966e0136c80036b277` |
+| Matching Stage1/Stage2 release | `phone20-stage12-v2` |
+| Reference source code | `https://github.com/ai4ce/dex_retarget` |
+| Collector reference commit | `30d43fba07a5a8bcbd756b1bf58be713daec5206` |
+| Proposed dataset | `SHARPA-BC-DATA-038 / phone8_mentor_tweak_recovery_v1` |
+| Dataset status | `DRAFT / ROLLOUTS_NOT_YET_MATERIALIZED` |
+
+Pin these identities. Do not silently replace a checkpoint, H5, GoalContract,
+Training Card, code commit, simulator runtime or collection schema.
+
+## What the repository contains
+
+```text
+objects/phone/releases/phone8-stage3-autodr-candidates-v1/
+  README.md
+  data_card.md
+  mentor_trajectory_generation_spec.md
+  release_manifest.json
+  checksums.sha256
+  episodes/<phone_id>/
+    checkpoint/model_<step>.pt
+    training_card.md
+    checkpoint_receipt.json
+    evaluation_summary.json
+
+objects/phone/releases/phone20-stage12-v2/
+  ... matching Stage1 NPZ, Stage2 H5, GoalContract, QA and provenance ...
+```
+
+The release provides checkpoint bytes and the information needed to build a
+versioned collection job. It does not provide newly generated augmentation
+rollouts, a prebuilt Isaac container, or a universally compatible one-command
+launcher.
+
+## Candidate inventory
+
+| Episode | Checkpoint | Published automatic evidence | Current boundary |
+|---|---|---|---|
+| `phone_02` | `model_7499.pt` | nominal seed0: 1/1 | nominal-only automatic evidence |
+| `phone_03` | `model_9500.pt` | not run | last complete checkpoint from crashed run |
+| `phone_05` | `model_7000.pt` | RefAug/reset DR seeds0-7: 6/8 | automatic failures remain |
+| `phone_07` | `model_7499.pt` | RefAug/reset DR seeds0-7: 8/8 | best starting point for the current collector family |
+| `phone_10` | `model_9000.pt` | not run | last complete checkpoint from crashed run |
+| `phone_14` | `model_8500.pt` | RefAug/reset DR seeds0-7: 7/8 | automatic failure remains |
+| `phone_17` | `model_7999.pt` | RefAug/reset DR seeds0-7: 8/8 | promising candidate; generic source adapter still required |
+| `phone_18` | `model_9998.pt` | nominal seed0: 1/1 | nominal-only evidence for this exact checkpoint |
+
+All eight remain non-promoted candidates. P2 robustness, automatic drop
+detection and human video review are incomplete.
+
+## Download and verify
+
+Install the Hugging Face CLI, then download the immutable releases:
+
+```bash
+python -m pip install "huggingface_hub[cli]"
+
+hf download FlappyHimself/sharpa-artifacts \
+  --repo-type dataset \
+  --revision 942f7cd025ae050823a817c9aff6cb7cb810d009 \
+  --include "objects/phone/releases/phone8-stage3-autodr-candidates-v1/**" \
+            "objects/phone/releases/phone20-stage12-v2/**" \
+  --local-dir ./sharpa-artifacts
+
+cd ./sharpa-artifacts/objects/phone/releases/phone8-stage3-autodr-candidates-v1
+sha256sum -c checksums.sha256
+```
+
+Clone and pin the reference implementation separately:
+
+```bash
+git clone https://github.com/ai4ce/dex_retarget.git
+cd dex_retarget
+git checkout 30d43fba07a5a8bcbd756b1bf58be713daec5206
+```
+
+Use the project runtime with Isaac Sim 5.1.0.0, Torch 2.7.0, `h5py` and
+`scipy`. The collector imports its H5 attempt-bank implementation even for
+`--help`, so a lightweight Python environment without `h5py/scipy` is not a
+valid collection environment.
+
+## Copy this task directly into Astra
+
+```text
+Use the public Hugging Face dataset FlappyHimself/sharpa-artifacts.
+
+Pin the teacher artifacts to revision
+942f7cd025ae050823a817c9aff6cb7cb810d009 and verify the release manifest and
+checksums before loading any checkpoint. Read, in order:
+
+1. objects/phone/releases/phone8-stage3-autodr-candidates-v1/README.md
+2. mentor_trajectory_generation_spec.md
+3. data_card.md
+4. the selected episode's training_card.md
+5. checkpoint_receipt.json
+6. evaluation_summary.json
+
+For one selected phone episode, resolve the matching Stage1 NPZ, Stage2 H5 and
+GoalContract from phone20-stage12-v2. Materialize a create-only collection spec,
+frozen attempt bank and runtime receipt. Run preflight before opening Isaac.
+
+Collect bounded tweak -> residual-teacher recovery -> functional-pose attempts.
+Persist the complete observation/action chain, actual simulator state,
+contact/penetration diagnostics, randomization ledger, metrics, provenance and
+canonical Stage1-MANO-over-policy-actual 2x3 video. Never use the raw residual as
+the BC label when the controller applies a processed/composed target. Preserve
+failed attempts as reflection/diagnostic data.
+
+Do not call a checkpoint robust or promoted. Do not put an attempt in the
+training dataset until its automatic gates and human video review pass. Never
+move a failed heldout attempt into train. Do not launch or resume teacher
+training as part of this collection task.
+```
+
+## Required augmentation loop
+
+For every attempt:
+
+1. Load exactly one episode's Training Card, checkpoint and matching reference
+   lineage.
+2. Reset to a preregistered phase and apply one bounded tweak ledger.
+3. Run a fresh Isaac/PhysX rollout through the complete reference plus a
+   15-control-step terminal hold.
+4. Record the causal transition:
+
+   ```text
+   obs_t
+     -> teacher raw residual_t
+     -> processed residual_t
+     -> reference/composed target_t
+     -> final controller target_t
+     -> four physics steps
+     -> actual state_t+1
+   ```
+
+5. Evaluate E4, E5, E6 and E9 and retain contact, penetration, drop/invalid and
+   termination diagnostics.
+6. Render the canonical 2x3 video: Stage1 MANO+object on the top row and the
+   policy's actual PhysX rollout on the bottom row, synchronized in front, side
+   and top views.
+7. Only an automatically valid and visually approved attempt may become the
+   parent of another bounded tweak. Parent and child remain immutable records.
+
+## Initial bounded tweak envelope
+
+Use the first release only inside the documented training-distribution envelope:
+
+| Tweak | Range |
+|---|---|
+| Common hand+phone X/Y translation | each axis `U[-0.05, 0.05] m` |
+| Common yaw | `U[-30, 30] deg` |
+| Phone-in-hand translation | each axis `U[-0.006, 0.006] m` |
+| Phone-in-hand axis-angle rotation | `U[-3, 3] deg` |
+| Positive finger wrap | `U[0, 0.06] rad` |
+| Four-finger rub | `U[-0.025, 0.025] rad` |
+
+Larger perturbations, pushes, physics randomization, observation noise or new
+tweak families require a new versioned case bank.
+
+## Attempt topology and outputs
+
+Freeze 128 attempts per independent teacher before observing results: 96 train
+and 32 heldout. Eight teachers define 1024 primary attempts. Train alone may be
+used for optimization and normalization; heldout is a one-shot
+same-distribution audit.
+
+Each create-only attempt must contain or link:
+
+```text
+attempt_<id>/
+  attempt_contract.json
+  metrics.json
+  rollout.npz
+  action_chain.npz
+  randomization_ledger.json
+  evaluation_suite_card.json
+  reference_vs_policy_2x3.mp4
+  render_manifest.json
+```
+
+Record checkpoint, Training Card, Stage1, Stage2, GoalContract, runtime, code,
+assets, seed, tweak, parent attempt and every delivered member by SHA256.
+
+## Current executor compatibility: read before running
+
+The public source tree contains the generic entry point
+`student/sharpa/teacher_rollout_collector.py`, whose command interface is:
+
+```bash
+COLLECTION_SPEC=/absolute/path/to/collection_spec.json
+ATTEMPT_BANK=/absolute/path/to/attempt_bank.json
+OUTPUT_ROOT=/absolute/path/to/new_collection_root
+
+python -m student.sharpa.teacher_rollout_collector \
+  --collection-spec "$COLLECTION_SPEC" \
+  --attempt-bank "$ATTEMPT_BANK" \
+  --row-start 0 \
+  --row-stop 1 \
+  --output-root "$OUTPUT_ROOT" \
+  --device cuda:0 \
+  --preflight-only
+```
+
+Do not run this command unchanged against all eight teachers. At the pinned
+reference commit, its v1 contract accepts only the older `phone07`/`phone18`
+source IDs, and its historical `71D -> 31D` student codec is not identical to
+the proposed `SHARPA-BC-DATA-038` schema in this release. The Hugging Face
+checkpoint bundle also does not include every runtime/environment receipt
+required by that collector's fail-closed preflight.
+
+Therefore:
+
+- use `phone_07` first when validating compatibility with the existing
+  collector family;
+- create a new versioned collection schema/adapter for the eight-teacher
+  campaign rather than weakening or bypassing preflight;
+- bind the selected Data Card's exact observation, action-chain and label
+  semantics;
+- run unit tests plus `--preflight-only` before any simulator rollout;
+- publish the materialized spec, attempt bank, runtime receipt and output
+  manifest under a new immutable dataset release.
+
+The release is directly usable as an artifact-and-contract handoff to Astra. It
+is not evidence that a portable eight-teacher collector has already passed
+preflight.
+
+## Completion criteria
+
+An Astra augmentation job is complete only when:
+
+- all inputs resolve to the pinned identities and pass hash checks;
+- every scheduled attempt has an immutable success, failure or infrastructure
+  record;
+- rollout/action timing and final applied targets are validated;
+- all accepted attempts pass the declared automatic gates and human review;
+- train/heldout membership remains frozen;
+- a final manifest, checksums and Data Card execution update are published.