跳转至

ReBRAC Phase 9 Implementation Plan

For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

Goal: Add a narrow but package-usable ReBRAC implementation for offline continuous-control training, plus the minimum offline dataset support needed to carry behavior actions for next states.

Architecture: Reuse the current TD3+BC offline runtime family instead of creating another trainer stack. Implement ReBRAC as a behavior-regularized TD3-style learner on top of MLPTD3Model, packaged configs, managed API wiring, and checkpoint/eval/predict paths, while extending TransitionDataset with optional next_actions so the learner can regularize bootstrapped targets without redesigning the whole data layer.

Tech Stack: Python, PyTorch, Gymnasium, existing rl_training offline dataset and experiment infrastructure

Task 1: Freeze The Narrow ReBRAC Scope

Files: - Create: docs/plans/2026-03-12-rebrac-phase9.md - Modify: README.md - Modify: docs/plans/2026-03-12-rl-expansion-roadmap-design.md - Modify: docs/plans/2026-03-12-rl-yearly-sourcebook-design.md

Step 1: Freeze v1 boundaries

Document the first packaged ReBRAC release as:

  • continuous Box actions only
  • flat vector observations only
  • offline dataset training only
  • single-process trainer only
  • TD3-style target smoothing with behavior regularization
  • no recurrent path
  • no image observations

Step 2: Record the package rationale

Explain that ReBRAC is the next low-friction 2023 offline wave because:

  • current public offline libraries still surface it
  • it complements the existing TD3+BC baseline
  • it reuses the current offline config / checkpoint / trainer stack

Step 3: Keep test execution deferred

Record that tests are added but intentionally not executed until the user explicitly requests it.

Task 2: Extend Offline Dataset Payloads For Next-State Behavior Actions

Files: - Modify: src/rl_training/data/offline_dataset.py - Modify: src/rl_training/data/dataset_loaders.py - Modify: src/rl_training/data/offline_mixers.py - Modify: src/rl_training/runtime/iql_trainer.py - Modify: src/rl_training/data/__init__.py - Modify: tests/test_offline_dataset.py - Modify: tests/test_dataset_loaders.py

Step 1: Add optional next_actions storage

Allow TransitionDataset to carry optional next_actions tensors without breaking existing callers that only expect the standard five transition fields.

Step 2: Preserve the field through loaders and mixing

When file payloads or Minari episodes provide next-state behavior actions, keep them through:

  • TransitionDataset.from_dict(...)
  • file-backed dataset loading
  • mixed dataset assembly
  • sampling to trainer batches

Step 3: Fill the field in random dataset generation

Generate sequential random offline datasets with behavior-consistent next_actions so package-provided smoke datasets already exercise the richer offline payload.

Task 3: Add The ReBRAC Learner

Files: - Create: src/rl_training/algorithms/rebrac.py - Modify: src/rl_training/algorithms/__init__.py

Step 1: Reuse the existing TD3 model family

Build ReBRAC on MLPTD3Model with target critics, target actor smoothing, and deterministic continuous actions.

Step 2: Implement behavior-regularized critic and actor objectives

Keep the package v1 knobs small and explicit:

  • actor_bc_weight
  • critic_bc_weight
  • actor_q_weight
  • policy_noise
  • noise_clip
  • policy_delay

Use sampled next_actions when present, and document the fallback behavior when a dataset batch does not carry them.

Step 3: Expose public loss helpers

Add a readable rebrac_loss(...) function and export ReBRAC / ReBRACAlgorithm through the shared algorithms package.

Task 4: Add The Offline ReBRAC Trainer

Files: - Create: src/rl_training/runtime/rebrac_trainer.py - Modify: src/rl_training/experiment/registry.py

Step 1: Reuse the offline dataset stack

Build the trainer on _infer_env_spaces(...) and _build_offline_dataset(...) from the current offline path.

Step 2: Preserve shared controls

Keep support for:

  • eval_interval
  • early stopping callbacks
  • offline epoch / update budgets
  • learning-rate schedules
  • checkpoint save / resume

Step 3: Reuse standard evaluation and prediction

Evaluate with the current deterministic continuous-action TD3 helper and expose package prediction through checkpoint workflows.

Task 5: Wire ReBRAC Into The Package Surface

Files: - Modify: src/rl_training/api/algorithms.py - Modify: src/rl_training/api/__init__.py - Modify: src/rl_training/__init__.py - Create: configs/rebrac/pendulum.yaml - Create: src/rl_training/assets/configs/rebrac/pendulum.yaml - Modify: README.md

Step 1: Add the managed API entrypoint

Expose ReBRAC through the root package and API namespaces.

Step 2: Ship a starter config

Add a packaged offline Pendulum-v1 config that uses the random dataset path and the narrow ReBRAC defaults.

Step 3: Update package docs

Document ReBRAC as a 2023 offline follow-on to TD3+BC and explain the optional next_actions dataset field.

Task 6: Add Unexecuted Coverage

Files: - Create: tests/test_rebrac_update.py - Create: tests/test_rebrac_trainer_smoke.py - Modify: tests/test_package_api_exports.py - Modify: tests/test_public_api.py - Modify: tests/test_experiment_manager.py - Modify: tests/test_checkpoint_workflows.py - Modify: tests/test_package_smoke.py - Modify: tests/test_cli.py

Step 1: Add learner-level coverage

Add a unit test for rebrac_loss(...) metric keys and one update call.

Step 2: Add trainer smoke coverage

Add a small offline smoke test that checks checkpoint creation and eval wiring.

Step 3: Extend package-surface expectations

Update public exports, managed API, checkpoint workflow, and packaged-config tests so ReBRAC is treated as a shipped algorithm.

Step 4: Keep execution deferred

Do not run the tests until the user explicitly asks for test execution.