---
pretty_name: Direction-Speed Reinforcement Learning Experiments
language:
  - en
  - zh
tags:
  - reinforcement-learning
  - robotics
  - experiment-results
  - action-space-shaping
---

# DS Experiments

Direction-Speed 动作分解实验数据集。对应代码仓库 [SZUEAILab/Q-chunking](https://github.com/SZUEAILab/Q-chunking)。

## 精简研究归档

新增 [2026-10-19 精简归档](summaries/2026-10-19/README.md)：包含历史分析图表、2026-10-18
研究报告、三组 RLPD 筛选的逐运行评估/配置、协议、来源与 SHA-256。保留负结果和删失边界，
不上传新权重、轨迹或逐步日志。文件一致性和运行完整性检查不等于科学结论通过。
详细清单见各归档的 `manifest.json`、`run_checks.json` 和 `SHA256SUMS`。

[2026-10-19 阶段总结](summaries/2026-10-19-phase-review/README.md)补充状态条件筛选的冻结数值分析、
失效退出凭证审计、Action Adapter 历史评估重核与 critic 机制诊断，并登记条件性后续计划。
旧报告不改判；72-run 新筛选的本地启动快照不代表本包已认证其完整终态。

## 结构

```
cube-triple-play-singletask/
└── {task}/
    └── {horizon}/
        └── {method}/
            └── {ds_mode}/
                └── seed{N}/
                    ├── eval.csv           # step, success, episode.return, ...
                    ├── flags.json         # ds_mode, horizon_length, seed, ...
                    ├── online_agent.csv   # critic_loss, actor_loss, alpha, ...
                    ├── env.csv            # qpos, qvel, time, ...
                    └── params_*.pkl       # 模型 checkpoint（仅 seed0）
```

## 目录

| 目录 | 内容 | Seeds | Eval | Runs |
|------|------|:---:|:---:|:---:|
| `task1/h1/rlpd/` | cube-triple-task1 RLPD H=1 | 4 | 50 | 16 |
| `task1/h5/rlpd/` | cube-triple-task1 RLPD-AC H=5 | 10 | 100 | 40 |
| `task1/h5/rlpd-4s/` | cube-triple-task1 RLPD-AC H=5（旧） | 4 | 50 | 16 |
| `task2/h1/rlpd/` | cube-triple-task2 RLPD H=1 | 4 | 50 | 16 |
| `task2/h5/rlpd/` | cube-triple-task2 RLPD-AC H=5 | 4 | 50 | 16 |
| `task2/h5/fql/` | cube-triple-task2 FQL H=5 | 3 | 50 | 6 |
| `task3/h1/rlpd/` | cube-triple-task3 RLPD H=1 | 10 | 100 | 40 |
| `task3/h5/rlpd/` | cube-triple-task3 RLPD-AC H=5 | 10 | 100 | 40 |
| `task4/h1/rlpd/` | cube-triple-task4 RLPD H=1 | 10 | 100 | 40 |
| `task4/h5/rlpd/` | cube-triple-task4 RLPD-AC H=5 | 10 | 100 | 40 |
| `task5/h1/rlpd/` | cube-triple-task5 RLPD H=1 | 10 | 100 | 40 |
| `task5/h5/rlpd/` | cube-triple-task5 RLPD-AC H=5 | 10 | 100 | 40 |

| 其他目录 | 内容 | 实验数 |
|------|------|:---:|
| `reproduce/` | QC 论文复现实验 | 4 |
| `parallel_benchmark/` | 并行训练压力测试 | 2 |

> **命名约定**：`h1`/`h5` = horizon length；`none` = TanhNormal baseline；`posthoc`/`stereographic`/`spherical` = DS 模式。`rlpd-4s` 为旧版 4-seed 实验，与新 10-seed 版区分。

## 使用

```python
from huggingface_hub import hf_hub_download

# 下载 task4 H=5 stereographic seed0 的 eval.csv
hf_hub_download("SZUEAILab/ds-experiments",
    "ds_experiments/cube-triple-play-singletask/task4/h5/rlpd/stereographic/seed0/eval.csv",
    repo_type="dataset")
```