[yuecideng]

Description

Policy behavior cloning and learned dynamics currently own separate optimizer and recovery loops. Add embodichain.learning.sl so structured inputs, objective-owned losses, single optimizer updates and offline training recovery can be shared without flattening model-specific state or changing loss reductions.

The layer provides owned tensor-tree sampling, explicit numerical/clipping contexts, source and semantic compatibility checks, atomic checkpoints, and the lazy embodichain train-sl recipe entry point. Sampling, normalization updates, physical acquisition and policy execution remain with their callers.

Dependencies

• Foundation layer 1/5; base: codex/nerd-model-data-training.
• Depends on Add NeRD model extension and an SL training showcase #855, which supplies the standalone NeRD baseline.
• Follow-up sibling layers: NeRD integration, offline visual BC, robot world models and RigidFormer adapters.
• Refs [Proposal] Unify differentiable environments and APG training infrastructure #848 and Add ODS visual BC learning and data-stream comparisons #828.
• Merge requires the requesting user's review. Auto-merge is not enabled.

Validation

• 19 focused SL tests cover independent masked SGD, multi-step gradients, finite-update boundaries, owned metadata/content identity, exact dropout/optimizer/sampler recovery, optimizer compatibility before mutation, and actual unified CLI routing.
• Existing context and API-checker suites pass; public API coverage 3128/3128.
• Relevant NeRD trainer coverage passes alongside the new test package.
• Black 26.10.1 black ., diff checks, context routing/checks pass.
• Sphinx dummy build succeeds with 982 existing site warnings; no global strict-build claim.
• Independent and automated review findings are fixed with regression coverage: owned metadata, optimizer/scheduler preflight, typed source identity, validation RNG/mode isolation, and preserving prior experiment artifacts, including delayed-fit output ownership.
• Context decision: add the requested SL owner, link learning-layer navigation, and preserve accurate RL/environment ownership.

Type of change

• New feature (non-breaking)

Checklist

• Run black . with Black 26.10.1.
• Review affected documentation and agent context.
• Document public APIs and pass the API coverage gate.
• Review test necessity and verify affected behavior.
• Preserve optional model/runtime dependency boundaries.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The reviewed changes appear safe to merge; no blocking issues remain.

Summary

Adds shared supervised-learning batches, optimizer updates, offline training, checkpoint recovery, and the lazy train-sl command.

• Supervised models keep their own loss rules while sharing one optimizer update.
• Offline training runs save their state so they can resume later.
• The main command line can start supervised training from a recipe.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Source["Source samples a batch"] --> Learner["Learner updates the model"]
  Learner --> Scheduler["Optional scheduler"]
  Scheduler --> Validation["Optional validation"]
  Validation --> Metrics["Write metrics"]
  Metrics --> Checkpoint["Save checkpoint"]
  Checkpoint --> Resume["Check compatibility and restore"]

Loading

_{Reviews (6) · Last reviewed commit: "fix(sl): recheck output ownership before..." · Reviewed by Greptile}