Skip to content

docs: add FlashSAC to support matrix and algorithms doc #340

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
TATP-233 merged 1 commit into from
Apr 24, 2026
Merged

docs: add FlashSAC to support matrix and algorithms doc #340

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions docs/users/zh_CN/02-simulation-backends.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,8 @@ uv run scripts/generate_support_matrix.py --write
| SAC (torch) | `g1_walk_rough` (G1 walk rough) | Tested | Registered |
| SAC (torch) | `g1_sac_wbt` (g1 sac wbt) | Tested | - |
| TD3 (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered |
| FlashSAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Tested |
| FlashSAC (torch) | `g1_walk_flat_amp` (g1 walk flat amp) | Tested | Registered |

### Source Index

Expand Down
35 changes: 35 additions & 0 deletions docs/users/zh_CN/04-algorithms.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -116,6 +116,41 @@ uv run scripts/train_offpolicy.py algo=td3 task=td3/g1_walk_flat/mujoco training
| `training.play_only` | false | 仅回放 |
| `training.no_play` | false | 跳过自动回放 |

## FlashSAC

FlashSAC 是基于 FlashAttention 风格 Block 的 off-policy 算法，actor 使用 BatchNorm embedder + 结构化噪声探索，critic 使用 distributional Q（C51 变体）。与 FastSAC 共用同一个训练入口 `train_offpolicy.py`，但网络结构和前向接口不同。

### Usage

```bash
# 基本训练
uv run scripts/train_offpolicy.py algo=flashsac task=flashsac/g1_walk_flat/mujoco

# 跳过自动回放
uv run scripts/train_offpolicy.py algo=flashsac task=flashsac/g1_walk_flat/mujoco training.no_play=true
```

### Playback

```bash
uv run scripts/train_offpolicy.py algo=flashsac task=flashsac/g1_walk_flat/mujoco training.play_only=true
uv run scripts/train_offpolicy.py algo=flashsac task=flashsac/g1_walk_flat/mujoco training.play_only=true algo.load_run="2026-04-23_14-06-57_mujoco"
```

### Key Parameters

| 参数 | 默认值 | 说明 |
|------|--------|------|
| `algo` | `flashsac` | 算法选择 |
| `task` | `flashsac/g1_walk_flat/mujoco` | 唯一 owner 配置入口；reward/env/algo 都在这里改 |
| `algo.max_iterations` | 5000 | 最大训练迭代数 |
| `algo.num_envs` | 1024 | 并行环境数量 |
| `algo.tau` | 0.01 | target network 软更新系数 |
| `algo.algo_params.actor_num_blocks` | 2 | actor FlashSAC block 层数 |
| `algo.algo_params.critic_num_blocks` | 2 | critic FlashSAC block 层数 |
| `training.play_only` | false | 仅回放 |
| `training.no_play` | false | 跳过自动回放 |

## Navigation

- Previous: [Training Guide](03-training.md)
Expand Down
7 changes: 7 additions & 0 deletions src/unilab/docs/support_matrix.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,13 @@ class SupportRow:
task_glob="*/*.yaml",
generic_tested=True,
),
EntrypointSpec(
entrypoint_id="flashsac_torch",
label="FlashSAC (torch)",
config_dir="conf/offpolicy/task/flashsac",
task_glob="*/*.yaml",
generic_tested=True,
),
)


Expand Down