[yuecideng]

Description

Add an engine-rendered RigidFormer sphere demo to the learning/rigidformer showcase, alongside pinned official-asset preparation, fixed-weight evaluation, raw/EMA export, independent author comparison and recovery checks.

embodichain run-example learning/rigidformer demo displays dynamic Newton XPBD spheres on the left and render-only model spheres on the right. The model receives four engine observations and performs one free rollout. The demo supports a native window or headless engine-camera recording, configurable resource paths and object selection, and writes MP4, a preview frame, aligned trajectories and source-identified metrics. Frames stream to the encoder. The README includes resource preparation and runnable commands.

Stack

• Layer: 3/3
• Base: codex/rigidformer-official-training
• Depends on: #889
• Followed by: none

Type of change

• New feature (non-breaking)

Validation

• Official Sphere weights and verified local MOVi-Sphere geometry: headless 8-frame smoke and default 51-frame collision/landing run; native-window path under Xvfb with six frames; two-object, five-frame smoke. All completed with cleanup.
• Default output verified as 960x540, 51 frames at 12 fps. Trajectory checks verify finite predictions, aligned shapes, observed warmup agreement and reported center errors.
• 39 existing showcase, example packaging and launcher tests pass. Demo help remains simulator-free; syntax checks pass. No unit tests of demo internals were added; the new example is exercised through the real CLI and engine.
• After synchronizing the Train official RigidFormer through shared SL #889 fixes, 20 focused model/SL checks pass (three GPU cases skipped in that CPU invocation). The inherited changes have their own CPU/Gloo/CUDA/NCCL validation in Train official RigidFormer through shared SL #889.
• Black 26.10.1 (black .), API coverage 3287/3287, context validation and 42 context tests pass. Neural-dynamics context and the example index were updated; simulator lifecycle and launcher guidance remain accurate.
• Pinned model parity, installed-wheel and fixed-weight evaluation evidence is recorded in the acceptance protocol and qualification index. The engine demo uses new initial conditions and Newton physics; its errors describe that visual comparison. Full published-benchmark evaluation and full-budget retraining remain separate.

Checklist

• Ran black . with Black 26.10.1.
• Reviewed affected documentation and agent context and updated the owning guidance.
• Public API documentation coverage passes.
• Verified proportional behavioral coverage and reviewed test necessity.
• The new example is preserved by the existing example packaging path; it uses existing engine and RigidFormer dependencies.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The changes add operator-run workflows around RigidFormer without changing general simulator behavior.

The PR appears safe to merge, with non-blocking improvements to raw-data checks and report reliability.

Findings

1.  Unlisted shards bypass checksum checks ▶
2.  Optional metadata can block reports ▶

Fix with agent prompt

### Issue 1
examples/learning/rigidformer/_prepare.py:126-133
`prepare` discovers every local Parquet shard but checks hashes only for paths listed in the official manifest. An extra shard in a reused cache can therefore supply scenes without an official checksum check. Its converted contents then receive the pinned `data_revision`.

Before creating the output directory, reject selected shards missing from the manifest or select shards from the manifest itself.

### Issue 2
examples/learning/rigidformer/_support/_evidence.py:66-69
If Torch supports NCCL but `libnccl.so.2` cannot be loaded by name, `evidence` raises `OSError`. The Torch build check does not guarantee that this separate lookup succeeds. Even a completed single-GPU evaluation then loses its summary.

Catch the lookup failure and record an unavailable runtime version instead of failing the experiment.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds the RigidFormer showcase with asset preparation, shared-training recipes, weight export, fixed-weight evaluation, author comparison and recovery checks.

• The RigidFormer showcase offers commands for its main workflows.
• A side-by-side demo shows Newton spheres and a RigidFormer rollout.
• Pinned MOVi assets become prepared data for RigidFormer workflows.
• Fixed-checkpoint evaluation reports RigidFormer results with their inputs.
• Comparison and recovery commands check RigidFormer training against references.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Pinned assets] --> B[Prepare verified geometry]
  B --> C[Shared SL training]
  C --> D[Export raw or EMA weights]
  D --> E[Fixed-weight evaluation]
  B --> F[Newton sphere demo]
  D --> F
  F --> G[Four observed frames]
  G --> H[One free model rollout]
  H --> I[Video and trajectory report]

Loading

_{Reviews (1) · Last reviewed commit: "style(examples): trim RigidFormer file e..." · Reviewed by Greptile}

[greptile-apps]

Unlisted shards bypass checksum checks

prepare discovers every local Parquet shard but checks hashes only for paths listed in the official manifest. An extra shard in a reused cache can therefore supply scenes without an official checksum check. Its converted contents then receive the pinned data_revision.

Before creating the output directory, reject selected shards missing from the manifest or select shards from the manifest itself.

Prompt To Fix With AI

This is a comment left during a code review.
Path: examples/learning/rigidformer/_prepare.py
Line: 126-133

Comment:
**Unlisted shards bypass checksum checks**

`prepare` discovers every local Parquet shard but checks hashes only for paths listed in the official manifest. An extra shard in a reused cache can therefore supply scenes without an official checksum check. Its converted contents then receive the pinned `data_revision`.

Before creating the output directory, reject selected shards missing from the manifest or select shards from the manifest itself.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Optional metadata can block reports

If Torch supports NCCL but libnccl.so.2 cannot be loaded by name, evidence raises OSError. The Torch build check does not guarantee that this separate lookup succeeds. Even a completed single-GPU evaluation then loses its summary.

Catch the lookup failure and record an unavailable runtime version instead of failing the experiment.

Prompt To Fix With AI

This is a comment left during a code review.
Path: examples/learning/rigidformer/_support/_evidence.py
Line: 66-69

Comment:
**Optional metadata can block reports**

If Torch supports NCCL but `libnccl.so.2` cannot be loaded by name, `evidence` raises `OSError`. The Torch build check does not guarantee that this separate lookup succeeds. Even a completed single-GPU evaluation then loses its summary.

Catch the lookup failure and record an unavailable runtime version instead of failing the experiment.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.