Repository navigation
Work: 稳定 FlashSAC 在 UniLab 中训练 Go2 & G1 并补充算法对比文档 #262
Description
Activity
Closing as completed.
Evidence:
- PR fix: Flashsac for g1_walk_flat #320 merged on 2026-04-20 and fixes FlashSAC G1 flat training stability, with successful ruff/mypy/pyright/test checks and training artifacts in the PR.
- PR tune flashsac g1_walk_flat training hyperparameters and fix flashsac … #336 merged on 2026-04-24 and fixes FlashSAC play/export behavior plus further G1 hyperparameter tuning.
- PR docs: add FlashSAC to support matrix and algorithms doc #340 merged on 2026-04-24 and adds FlashSAC support-matrix / algorithm docs.
- PR add Go2JoystickFlat task owner and retire G1_walk_flat_amp task #351 merged on 2026-04-29 and adds the Go2JoystickFlat FlashSAC task owner with training artifacts.
- The issue's Definition of Done is already checked off in the issue body.
Note: some later PR checks are marked skipped by workflow policy, and the attached artifacts are TensorBoard / GitHub attachments rather than W&B links. Treating those as acceptable project evidence, the remaining work is not repo-blocking for this issue.
Work type
benchmark
Area
G1/GO2
Problem
FlashSAC 已完成迁移至 UniLab，但当前训练过程不稳定，在 locomotion 任务上存在训练发散或无法稳定收敛的问题。在没有可靠基线的情况下，无法对 FlashSAC 与其他 off-policy 算法进行有效对比，也无法将其作为后续工作的参考实现。此外，目前缺少一份记录各算法对比结果的文档，导致贡献者难以了解 UniLab 中各算法的整体表现。
Deliverable
docs/benchmarks/algorithm_comparison.md），记录 PPO、FastSAC、FlashSAC 等已支持算法在 Go2 和 G1 上的benchmark 结果
Definition of done
docs/目录