Skip to content

refactor: unify g1 walk env tasks - #322

Merged
TATP-233 merged 5 commits into
mainfrom
fix/issue-175-g1-single-env-class
Apr 22, 2026
Merged

TATP-233 merged 5 commits into
mainfrom
fix/issue-175-g1-single-env-class

Conversation

@TATP-233

Copy link
Copy Markdown
Collaborator

Fixes #175

Summary

  • merge the G1 walk SAC/PPO env implementations into a single G1WalkEnv class
  • move task differentiation to owner YAML/config while keeping effective training behavior unchanged
  • unify G1 task names to g1_walk_xxx and update related tests, docs, and benchmarks
  • remove invalid critic-noise config paths and hardcode observation scaling in Python

Validation

  • make test
  • pre-commit on commit (ruff, ruff format, pyright, mypy)

@TATP-233
TATP-233 force-pushed the fix/issue-175-g1-single-env-class branch from 2318a5e to 7b1c8e1 Compare April 22, 2026 02:44
- split_obs_dict / get_obs_dims: when no critic group, critic = obs
  (never return None or 0)
- All learners (TD3/SAC/FlashSAC): critic_obs_dim is now a required
  parameter, no default-0-then-fallback pattern
- update_critic / update_actor: use data["critic"] directly instead
  of data.get("critic") with fallback to actor obs
- OffPolicyRunner: rename critic_dim → critic_obs_dim for consistency
- Collectors: critic_np is always a concrete array, no None branches
- FastTD3Runner: pass env_cfg_override to get_env_dims and
  critic_obs_dim to learner (fixes td3/g1_walk_flat config load)
@TATP-233
TATP-233 merged commit 0edb4d3 into main Apr 22, 2026
10 checks passed
@TATP-233
TATP-233 deleted the fix/issue-175-g1-single-env-class branch April 22, 2026 09:07
TATP-233 added a commit that referenced this pull request May 4, 2026
TATP-233 added a commit that referenced this pull request May 30, 2026
TATP-233 added a commit that referenced this pull request Sep 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Work: Reward 配置动态扩展机制 — 消除硬编码奖励类型

1 participant