|
Download README.md from yangfy0627/RL-postTrain: direct link, hf CLI and curl.
- Browser
- Download file 4.23 kB
-
https://huggingface.co/yangfy0627/RL-postTrain/resolve/main/README.md
- Command line
-
hf download hf://yangfy0627/RL-postTrain/README.md
-
curl -L -o README.md https://huggingface.co/yangfy0627/RL-postTrain/resolve/main/README.md
4.23 kB
| language: | |
| - zh | |
| tags: | |
| - robotics | |
| - reinforcement-learning | |
| - behavior-cloning | |
| - robomimic | |
| # RL-postTrain | |
| Robomimic 的 study1 与 study2 实验产物仓库,以及 DexMimicGen 的 study1 BC。 | |
| ```text | |
| Robomimic/ | |
| ├── study1/ | |
| │ ├── Transport/ | |
| │ ├── ToolHang/ | |
| │ ├── Square/ | |
| │ └── Can/ | |
| └── study2/ | |
| ├── Dice/ | |
| │ ├── step1-vanilla/ | |
| │ ├── step2-success_loss/ | |
| │ └── step2-critic_ca/ | |
| ├── OGPO/ | |
| │ ├── step1-vanilla/ | |
| │ ├── step2-success_loss/ | |
| │ └── step2-critic_ca/ | |
| ├── QC/ | |
| │ ├── step1-vanilla/ | |
| │ ├── step2-success_loss/ | |
| │ └── step2-critic_ca/ | |
| └── DSRL/ | |
| ├── step1-vanilla/ | |
| ├── step2-success_loss/ | |
| └── step2-critic_ca/ | |
| ``` | |
| ```text | |
| DexMimicGen/ | |
| └── study1/ | |
| ├── Threading/ | |
| ├── DrawerCleanup/ | |
| └── LiftTray/ | |
| ``` | |
| 目录按 study1 / study2 组织。已上传的 BC 权重见下方索引。 | |
| <!-- completed-bc:start --> | |
| ## 已完成的 BC | |
| | 任务 | 数据组 | Seed | Epoch | 最终50局成功率 | | |
| |---|---|---|---|---| | |
| | ToolHang | [diverse50](Robomimic/study1/ToolHang/diverse50/seed_1/) | 1 | 3000 | 6/50 = 12% | | |
| | ToolHang | [focused50](Robomimic/study1/ToolHang/focused50/seed_1/) | 1 | 3000 | 6/50 = 12% | | |
| | ToolHang | [mixed50](Robomimic/study1/ToolHang/mixed50/seed_1/) | 1 | 3000 | 5/50 = 10% | | |
| | Transport | [diverse50](Robomimic/study1/Transport/diverse50/seed_1/) | 1 | 3000 | 30/50 = 60% | | |
| | Transport | [focused50](Robomimic/study1/Transport/focused50/seed_1/) | 1 | 3000 | 32/50 = 64% | | |
| | Transport | [mixed50](Robomimic/study1/Transport/mixed50/seed_1/) | 1 | 3000 | 22/50 = 44% | | |
| 完整配置和记录见 [study1 索引](Robomimic/study1/README.md)。 | |
| <!-- completed-bc:end --> | |
| <!-- dmg-best-bc:start --> | |
| ## DexMimicGen BC 最佳保存点 | |
| | 任务 / 数据组 | 最佳 epoch | 窗口成功率(150局) | 该 epoch 单次(50局) | 选取时进度 | | |
| |---|---|---|---|---| | |
| | Threading [diverse100](DexMimicGen/study1/Threading/diverse100/seed_1/) | 1650 | 37.3% | 38% | 2045/3000(用户停止) | | |
| | Threading [mixed100](DexMimicGen/study1/Threading/mixed100/seed_1/) | 1400 | 34.0% | 无(该 epoch 未评估) | 2041/3000(用户停止) | | |
| | DrawerCleanup [diverse50](DexMimicGen/study1/DrawerCleanup/diverse50/seed_1/) | 1750 | 68.7% | 无(该 epoch 未评估) | 1828/3000(用户停止) | | |
| | DrawerCleanup [mixed50](DexMimicGen/study1/DrawerCleanup/mixed50/seed_1/) | 1150 | 72.0% | 无(该 epoch 未评估) | 1830/3000(用户停止) | | |
| | LiftTray [diverse50](DexMimicGen/study1/LiftTray/diverse50/seed_1/) | 700 | 49.3% | 无(该 epoch 未评估) | 1180/3000(用户停止) | | |
| | LiftTray [mixed50](DexMimicGen/study1/LiftTray/mixed50/seed_1/) | 850 | 40.7% | 无(该 epoch 未评估) | 1263/3000(用户停止) | | |
| 见 [DexMimicGen/study1](DexMimicGen/study1/README.md)。 | |
| <!-- dmg-best-bc:end --> | |
| <!-- visual-no-collision-bc:start --> | |
| ## 纯视觉输入 BC(不渲染碰撞网格) | |
| 当前已定稿 9 项;本机训练的 9 项按用户决定(2026-10-27)统一使用 epoch 1400。状态、权重与 200 局评估见 [visual_no_collision](visual_no_collision/README.md)。请按 selected_after_joint_review 标记识别定稿权重;定稿不表示零过拟合或全局最优。旧版本均在各自 releases/ 下保留。 | |
| <!-- visual-no-collision-bc:end --> | |
| <!-- full-rl-artifacts:start --> | |
| ## RL 完整产物 | |
| 2026-10-29 批次:已上传并核验 20/20 个完整 run 归档;ToolHang50 不在本次范围内。所有已归档 run 均有真实 300 局终评,其中 3 个按零成功率规则提前结束,已保留状态标记。 | |
| 完整保留的 run 目录包括 checkpoint、replay、offline buffer、配置、日志、评估、诊断和 W&B 文件,并附 BC、源码、数据清单、逐文件校验和恢复脚本。原始示范数据及渲染图像输入保留哈希引用。 | |
| 查看 [ToolHang100](RL/ToolHang100/README.md)、[Transport50](RL/Transport50/README.md) 和 [本批完整清单](RL/full_artifacts_20261029.json)。 | |
| <!-- full-rl-artifacts:end --> | |