Feature/flashsac integration - #227
Open
psurchit wants to merge 6 commits into
Open
Conversation
Port FlashSAC's weight-normalized layers, squashed-Gaussian policy, and ensembled categorical (distributional) double-critic into rsl_rl as modules/flashsac_layers.py, and wrap them as RSL-RL models (FlashSACActorModel, FlashSACCriticModel) following the MLPModel surface (obs-group resolution, EmpiricalNormalization, JIT/ONNX export). Actor exports a deterministic tanh(mean) head; ONNX export verified. Dual BSD-3 / MIT (FlashSAC) attribution. Passes ruff check + format.
Cover actor/critic shapes and action bounds, distributional prob-sum, weight-normalization unit rows, tanh log-det-Jacobian parity, temperature init, obs-group fail-loud (KeyError/ValueError), and ONNX export validity + eager parity. 16 tests, mirroring the existing tests/ layout.
Fixed-capacity uniform replay buffer storing TensorDict observation groups, with n-step return accumulation ported from FlashSAC's TorchUniformBuffer. Provides add/can_sample/sample (Batch with terminated/truncated/next_obs) and save/load. 9 unit tests cover n-step discounting, done-truncation, wrap-around, sampling shapes, fail-loud construction, and persistence. Passes ruff + pyright; full suite 200 passed.
Add the FlashSAC algorithm implementing the RSL-RL algorithm contract (act/process_env_step/compute_returns no-op/update/save/load/get_policy/ compile/construct_algorithm): squashed-Gaussian actor, ensembled categorical double critic with EMA target, auto-tuned entropy temperature, n-step categorical TD targets, delayed actor updates, and zeta noise-repetition exploration. Eager-only + single-GPU for v1 (rejects use_amp/compile/multi-GPU loudly). Reconstructs terminated/truncated from dones + time_outs, requiring time_outs for infinite-horizon envs (fail loud). Adds utils: warmup_cosine_decay scheduler, RewardNormalizer, and fail-loud _require + resolve_sac_obs_groups helpers. 17 tests cover construction, rollout/exploration, one update step, delayed actor, EMA drift, save/load, and fail-loud config paths. ruff + pyright clean; full suite 217 passed.
Add OffPolicyRunner (subclass of OnPolicyRunner) that overrides learn() with
an off-policy loop: collect num_steps_per_env interactions into the replay
buffer, then run num_steps_per_env * updates_per_step gradient updates. Reuses
the base construction, logging, checkpointing, and ONNX/JIT export. Rejects
multi-GPU (WORLD_SIZE > 1) at construction.
Also make the shared Logger robust to algorithms without RND: read
cfg['algorithm'].get('rnd_cfg') instead of requiring the key (no behavior
change for PPO). 10 runner tests cover construction, multi-GPU rejection,
learn loop, param updates, save/load (incl. actor-only), inference bounds,
and ONNX export. ruff + pyright clean; full suite 227 passed.
Average each metric over the iteration's gradient updates instead of keeping only the last update's dict. With delayed actor updates (actor_update_period > 1) and updates_per_step making the last update a critic-only step, actor/entropy/temperature metrics were silently dropped from logging even though the actor was training. Now all reported metrics surface.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integrated FlashSAC (paper, code) - published at RSS 2026. I will also be raising a PR on MJLab to integrate the MJLab wrapper corresponding to this.