[yuecideng]

Description

Route standalone NeRD optimizer updates through learning.sl while retaining the #855 codec, numerical profiles, normalization, learning-rate timing, legacy checkpoints and portable exports. Add Cartpole/Ant train-sl recipes that use the common offline trainer, plus an independent reference/candidate qualification runner.

Dependencies

• Model layer2/5; base: codex/sl-foundation; depends on Add shared SL updates, offline training and recovery #857 and Add NeRD model extension and an SL training showcase #855.
• Reference Add NeRD model extension and an SL training showcase #855 is pinned at a58de92a82b4e1c1c260083b5b1930173322aceb; custom neural-solver deployment remains with Add neural dynamics solver and differentiable Scene integration #822.
• Independent of the BC, RWM and RigidFormer sibling PRs.
• Merge requires the requesting user's review. Auto-merge is not enabled.

Validation

• Full clean-source GPU1 qualification passes26/26 gates: Cartpole and Ant ×5 paired seeds ×1000 updates, complete published held-out test evaluation, first-update forward/raw-gradient/clipped-gradient/parameter capture, final weights and restarted500+500 recovery.
• First-update tensors, final weights and resumed training states match Add NeRD model extension and an SL training showcase #855 exactly. Both paired one-sided95% degradation upper bounds are0. CPU model/codec/window/trainer/CLI coverage also passes; optional reference/resource cases remain explicitly skipped.
• Published model/data revisions, complete metrics, dependency/source hashes, phase times and CPU/Torch-GPU memory scopes are frozen in /data1/dyc/.codex-artifacts/embodichain-sl-nerd-20261025/final/{summary.json,report.md}. Reproduction code and protocol live in scripts/experiments/supervised/nerd*.
• The final numerical matrix uses candidate commit00b6a56a0. Subsequent changes only add API navigation and shared metadata/recovery preflight checks; focused integration/recovery tests pass after those changes.
• Black26.10.1 black ., API3132/3132, context and diff checks pass.
• This qualifies standalone transition fitting and compatibility. Historical Scene-based long-rollout numbers are not applied to the direct-Newton physical domain.
• Context decision: revise the NeRD owner to document shared updates and both preserved legacy/new training entry points.

Qualification admission follow-up

Commitd591e78ba rejects duplicate numeric or malformed seeds before imports/output/workers, while preserving short unique diagnostic runs as pending. Seven focused parser/CLI cases and the existing trainer tests pass; the completed five-independent-seed numerical matrix is unchanged.

Type of change

• Enhancement (non-breaking)

Checklist

• Run black . with Black26.10.1.
• Review affected docs and project context.
• Document public APIs and pass API coverage.
• Reuse existing coverage and add meaningful shared-update/recipe checks.
• Preserve dependencies and simulator-free training.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the previous repeated-seed issue is fixed and no new blocking issue was found.

Summary

NeRD training now uses the shared supervised learner while keeping its existing training and checkpoint behavior.

• NeRD’s existing trainer now uses the shared learner for each update.
• Cartpole and Ant NeRD runs can use the common train-sl trainer.
• A paired runner compares reference and candidate NeRD training runs.

The previous repeated-seed finding is fully addressed. No new actionable issues were found.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["Existing dynamics workflow"] --> B["DynamicsTrainer"]
  C["train-sl recipe"] --> D["SupervisedTrainer"]
  B --> E["SupervisedLearner"]
  D --> E
  E --> F["NeRDObjective"]
  F --> G["NeRD model and codec"]
  B --> H["Legacy checkpoints"]
  D --> I["SL checkpoints"]

Loading

_{Reviews (2) · Last reviewed commit: "fix(sl): reject duplicate NeRD acceptanc..." · Reviewed by Greptile}