chore: 移除仓内 T800 任务，完成到 engineai_rl_unilab 的外迁 (#1490)

.gitignore

@@ -82,10 +82,6 @@ src/unilab/assets/motions/x2/*.csv
 src/unilab/assets/robots/x2/meshes/*.STL
 src/unilab/assets/robots/microduck/assets/*
 !src/unilab/assets/robots/microduck/assets/.gitkeep
-src/unilab/assets/robots/t800/assets/*
-!src/unilab/assets/robots/t800/assets/.gitkeep
-src/unilab/assets/robots/t800/textures/*
-!src/unilab/assets/robots/t800/textures/.gitkeep
 src/unilab/assets/robots/g1/assets/
 src/unilab/assets/robots/g1/textures/
 src/unilab/assets/robots/go2/assets/

docs/sphinx/source/en/4-developer_guide/7-motion_assets.md

@@ -88,7 +88,7 @@ Robot binary meshes and textures (for example `.STL`, `.obj`, and `.png`) are
 externalized the same way, on the Hugging Face dataset repo
 [unilabsim/unilab-robots](https://huggingface.co/datasets/unilabsim/unilab-robots).
 The registered robots are a2, allegro_hand, g1, go2, go2_arm, microduck,
-sharpa_wave, t800, and x2 (`ROBOT_ASSET_SPECS` in `src/unilab/assets/hub.py`).
+sharpa_wave, and x2 (`ROBOT_ASSET_SPECS` in `src/unilab/assets/hub.py`).
 Their mesh/texture directories download lazily on first use and land under
 their original paths (for example `src/unilab/assets/robots/g1/assets/` and
 `robots/g1/textures/` for G1), so the original relative XML paths remain
@@ -107,30 +107,29 @@ To add a new robot's binary assets:
 
 1. Upload each directory to the HF repo while keeping the directory layout
    identical. A robot with multiple asset directories requires one upload per
-   directory. For example, T800 uses:
+   directory. For example, G1 uses:
 
    ```bash
    uv run hf upload unilabsim/unilab-robots \
-     src/unilab/assets/robots/t800/assets robots/t800/assets \
+     src/unilab/assets/robots/g1/assets robots/g1/assets \
      --repo-type dataset
    uv run hf upload unilabsim/unilab-robots \
-     src/unilab/assets/robots/t800/textures robots/t800/textures \
+     src/unilab/assets/robots/g1/textures robots/g1/textures \
      --repo-type dataset
    ```
 
 2. Ignore the downloaded directory contents in `.gitignore` (robots whose
    entire directory is HF-hosted ignore the whole directory; the older
-   microduck/t800 entries keep a `.gitkeep`), exclude the directory in
+   microduck entry keeps a `.gitkeep`), exclude the directory in
    `tool.uv.build-backend.source-exclude`, and register it in
    `ROBOT_ASSET_SPECS`.
 3. Scenes built through `create_backend` are then covered automatically:
    `ensure_robot_assets_for_paths` resolves the registered directories on a
    cold path before any backend parses the XML. Entry points that bypass
-   `create_backend` resolve explicitly, e.g. a T800 task factory calls:
+   `create_backend` resolve explicitly, e.g. the MicroDuck task factory calls:
 
    ```python
-   resolve_robot_asset_dir("robots/t800/assets", marker="LINK_BASE.obj")
-   resolve_robot_asset_dir("robots/t800/textures", marker="LINK_BASE.png")
+   resolve_robot_asset_dir("robots/microduck/assets", marker="trunk_base.stl")
    ```
 
 ## Architecture Notes
@@ -149,6 +148,6 @@ To add a new robot's binary assets:
   original local path exactly.
 - Robot binary assets use the same directory resolver
   (`resolve_robot_asset_dir`). The
-  thin X2, MicroDuck, and T800 task factories resolve their directories once
+  thin X2 and MicroDuck task factories resolve their directories once
   before delegating to the shared manager environment factory. The resolver is
   also exposed through the `unilab-pull-assets` CLI.

docs/sphinx/source/en/5-reference/5-support_matrix.md

@@ -13,9 +13,9 @@ in English. Do not infer support beyond the evidence grade shown below.
   headless physics in an external Python 3.11 worker; its owner scope is
   currently limited to the configured G1 walk-flat PPO/SAC paths.
 - `--sim mjwarp` has completed training validation only on the `g1_walk_flat`
-  host adapter, where PPO (torch) and SAC (torch) are Tested; the SAC
-  `t800_walk_flat` mjwarp owner is Configured only, other entrypoints follow
-  the matrix below, and using it requires installing the `mjwarp` extra.
+  host adapter, where PPO (torch) and SAC (torch) are Tested; other
+  entrypoints follow the matrix below, and using it requires installing the
+  `mjwarp` extra.
 - `--algo`, `--task`, and `--sim` jointly select the owner YAML.
 - Do not treat `training.sim_backend` as a standalone backend switch.
 
@@ -53,9 +53,7 @@ owner may only cover training smoke and an explicitly enabled DR subset.
 `mjwarp` has completed training validation only on the `g1_walk_flat` host
 adapter: the PPO (torch) and SAC (torch) owners have completed training
 validation and have backend, contract, and playback automated coverage, so
-they are marked `Tested`. The SAC `t800_walk_flat` mjwarp owner only has an
-owner YAML and compose coverage, so it is marked `Configured`, which does not
-imply training validation. mjwarp playback
+they are marked `Tested`. mjwarp playback
 supports explicit, finite-step `record` and reuses the MuJoCo offline
 renderer; `uv run eval --sim mjwarp --render-mode interactive` routes to the
 MuJoCo interactive viewer (mjwarp physics, MuJoCo rendering of env[0], forced
@@ -110,7 +108,6 @@ rendering/playback paths remain unsupported.
 | PPO (torch) | `go2w_joystick_flat` (go2w joystick flat) | Tested | - | Tested | - | - |
 | PPO (torch) | `go2w_joystick_rough` (go2w joystick rough) | Tested | - | Tested | - | - |
 | PPO (torch) | `stewart_balance` (stewart balance) | Tested | - | Tested | - | - |
-| PPO (torch) | `t800_walk_flat` (t800 walk flat) | Tested | Registered | - | - | - |
 | APPO (torch) | `go1_joystick_flat` (Go1 joystick) | Tested | - | Tested | - | - |
 | APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Tested | - | - |
 | APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Registered | Registered | Registered |
@@ -137,7 +134,6 @@ rendering/playback paths remain unsupported.
 | SAC (torch) | `g1_23dof_wall_flip_tracking` (g1 23dof wall flip tracking) | Tested | - | Registered | - | - |
 | SAC (torch) | `g1_23dof_wbt_obs` (g1 23dof wbt obs) | Tested | - | Registered | - | - |
 | SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | Registered | - | - |
-| SAC (torch) | `t800_walk_flat` (t800 walk flat) | Tested | Configured | - | - | - |
 | TD3 (torch) | `go1_joystick_flat` (Go1 joystick) | Registered | - | Tested | - | - |
 | TD3 (torch) | `go2_joystick_flat` (Go2 joystick) | Registered | - | Tested | - | - |
 | TD3 (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Registered | Registered | Registered |

docs/sphinx/source/zh_CN/4-developer_guide/7-motion_assets.md

@@ -81,7 +81,7 @@ env:
 托管在 Hugging Face 数据集仓库
 [unilabsim/unilab-robots](https://huggingface.co/datasets/unilabsim/unilab-robots)。
 已注册的机器人为 a2、allegro_hand、g1、go2、go2_arm、microduck、sharpa_wave、
-t800、x2（见 `src/unilab/assets/hub.py` 的 `ROBOT_ASSET_SPECS`）。它们的
+x2（见 `src/unilab/assets/hub.py` 的 `ROBOT_ASSET_SPECS`）。它们的
 mesh/纹理目录在首次使用时按需下载，落盘到原始路径（例如 G1 的
 `src/unilab/assets/robots/g1/assets/` 与 `robots/g1/textures/`），因此 XML 中的
 原始相对路径保持有效。这些目录通过 `pyproject.toml` 的
@@ -97,28 +97,27 @@ uv run unilab-pull-assets --robot all   # 所有已注册机器人
 新增某个机器人的二进制资产：
 
 1. 按目录上传到 HF 仓库，保持目录结构一致。一个机器人有多个资产目录时，
-   每个目录分别上传。例如 T800：
+   每个目录分别上传。例如 G1：
 
    ```bash
    uv run hf upload unilabsim/unilab-robots \
-     src/unilab/assets/robots/t800/assets robots/t800/assets \
+     src/unilab/assets/robots/g1/assets robots/g1/assets \
      --repo-type dataset
    uv run hf upload unilabsim/unilab-robots \
-     src/unilab/assets/robots/t800/textures robots/t800/textures \
+     src/unilab/assets/robots/g1/textures robots/g1/textures \
      --repo-type dataset
    ```
 
 2. 在 `.gitignore` 中忽略下载目录（整个目录托管到 HF 的机器人直接忽略整个目录；
-   较早的 microduck/t800 条目保留 `.gitkeep`），在
+   较早的 microduck 条目保留 `.gitkeep`），在
    `tool.uv.build-backend.source-exclude` 中排除该目录，并在 `ROBOT_ASSET_SPECS`
    中注册。
 3. 经由 `create_backend` 构建的 scene 会被自动覆盖：
    `ensure_robot_assets_for_paths` 在 backend 解析 XML 之前的冷路径解析已注册目录。
-   绕过 `create_backend` 的入口需要显式解析，例如 T800 task factory 调用两次：
+   绕过 `create_backend` 的入口需要显式解析，例如 MicroDuck task factory：
 
    ```python
-   resolve_robot_asset_dir("robots/t800/assets", marker="LINK_BASE.obj")
-   resolve_robot_asset_dir("robots/t800/textures", marker="LINK_BASE.png")
+   resolve_robot_asset_dir("robots/microduck/assets", marker="trunk_base.stl")
    ```
 
 ## 架构说明
@@ -131,6 +130,6 @@ uv run unilab-pull-assets --robot all   # 所有已注册机器人
   `ensure_robot_assets_for_paths`，然后再分发给具体 backend。
 - 热路径（`step` / `reset`）**不会**触发任何文件下载或解析。
 - `ASSETS_ROOT_PATH` 定义不变，下载落盘位置与原始本地路径完全一致。
-- 机器人二进制资产使用同一目录 resolver（`resolve_robot_asset_dir`）。X2、
-  MicroDuck 和 T800 的薄 task factory 都会先在冷路径解析一次，再委托给共享
+- 机器人二进制资产使用同一目录 resolver（`resolve_robot_asset_dir`）。X2 和
+  MicroDuck 的薄 task factory 都会先在冷路径解析一次，再委托给共享
   manager env factory；同一 resolver 也通过 `unilab-pull-assets` CLI 暴露。

docs/sphinx/source/zh_CN/5-reference/5-support_matrix.md

@@ -46,7 +46,7 @@ uv run scripts/generate_support_matrix.py --write
 
 `Tested` 只描述仓库中已有自动化覆盖或显式 maintainer 训练验证，不代表该组合具备同名 MuJoCo owner 的全部 backend capability；例如 phase-1 Motrix owner 可能只覆盖训练 smoke 和明确启用的 DR 子集。
 
-`mjwarp` 完成训练验证的只有 `g1_walk_flat` host adapter：PPO (torch) 与 SAC (torch) owner 已完成训练验证，并有 backend、contract 与 playback 自动化覆盖，因此标记为 `Tested`。SAC `t800_walk_flat` 的 mjwarp owner 只有 owner YAML 与 compose 覆盖，标记为 `Configured`，不代表训练验证。mjwarp playback 默认仅支持显式、有限步数的 `record` 并复用 MuJoCo 离线 renderer；`uv run eval --sim mjwarp --render-mode interactive` 路由到 MuJoCo 交互 viewer（mjwarp 跑物理、MuJoCo 渲染 env[0]，强制单 env）；不支持 `auto` 或 native playback。其他 entrypoint 中出现的 `Registered` 只表示 env/backend registry identity，不代表对应算法、terrain、完整 DR 或 production training 支持。
+`mjwarp` 完成训练验证的只有 `g1_walk_flat` host adapter：PPO (torch) 与 SAC (torch) owner 已完成训练验证，并有 backend、contract 与 playback 自动化覆盖，因此标记为 `Tested`。mjwarp playback 默认仅支持显式、有限步数的 `record` 并复用 MuJoCo 离线 renderer；`uv run eval --sim mjwarp --render-mode interactive` 路由到 MuJoCo 交互 viewer（mjwarp 跑物理、MuJoCo 渲染 env[0]，强制单 env）；不支持 `auto` 或 native playback。其他 entrypoint 中出现的 `Registered` 只表示 env/backend registry identity，不代表对应算法、terrain、完整 DR 或 production training 支持。
 
 `isaacgym` 是 Python 3.8 子进程后端，当前只接入 `g1_walk_flat`。SAC (torch) owner 已在真机（external Python 3.8 worker runtime，不在仓库 CI 覆盖）完成训练与 record playback 验证，标记为 `Tested`；其余 isaacgym cell 最高只到 `Configured`（registry + owner YAML + compose/contract 覆盖），不代表任何训练或 play 验证。playback 走 IsaacGym 原生渲染（viewer + camera sensor 离屏录制），有显示器时 `play_render_mode=auto` 打开交互 viewer，无显示器时自动降级为离屏录制。
 
@@ -95,7 +95,6 @@ uv run scripts/generate_support_matrix.py --write
 | PPO (torch) | `microduck_velocity_bam_flat` (microduck velocity bam flat) | Tested | - | - | - | - | - |
 | PPO (torch) | `microduck_velocity_flat` (microduck velocity flat) | Tested | Configured | - | - | - | - |
 | PPO (torch) | `stewart_balance` (stewart balance) | Tested | - | Tested | - | - | - |
-| PPO (torch) | `t800_walk_flat` (t800 walk flat) | Tested | Registered | - | - | - | - |
 | APPO (torch) | `go1_joystick_flat` (Go1 joystick) | Tested | - | Tested | - | - | - |
 | APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Tested | - | - | - |
 | APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Registered | Registered | Registered | Registered |
@@ -123,7 +122,6 @@ uv run scripts/generate_support_matrix.py --write
 | SAC (torch) | `g1_23dof_wbt_obs` (g1 23dof wbt obs) | Tested | - | Registered | - | - | - |
 | SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | Registered | - | - | - |
 | SAC (torch) | `microduck_velocity_flat` (microduck velocity flat) | Tested | Configured | - | - | - | - |
-| SAC (torch) | `t800_walk_flat` (t800 walk flat) | Tested | Configured | - | - | - | - |
 | TD3 (torch) | `go1_joystick_flat` (Go1 joystick) | Registered | - | Tested | - | - | - |
 | TD3 (torch) | `go2_joystick_flat` (Go2 joystick) | Registered | - | Tested | - | - | - |
 | TD3 (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Registered | Registered | Registered | Registered |

pyproject.toml

@@ -17,8 +17,6 @@ source-exclude = [
     "/src/unilab/assets/robots/go2_arm/assets",
     "/src/unilab/assets/robots/microduck/assets",
     "/src/unilab/assets/robots/sharpa_wave/meshes",
-    "/src/unilab/assets/robots/t800/assets",
-    "/src/unilab/assets/robots/t800/textures",
     "/src/unilab/assets/robots/x2/meshes",
 ]
 

scripts/tools/support_matrix.py

@@ -338,8 +338,6 @@ def render_support_matrix(root: Path | None = None) -> str:
         "",
         "`mjwarp` 完成训练验证的只有 `g1_walk_flat` host adapter：PPO (torch) 与 SAC (torch) owner "
         "已完成训练验证，并有 backend、contract 与 playback 自动化覆盖，因此标记为 `Tested`。"
-        "SAC `t800_walk_flat` 的 mjwarp owner 只有 owner YAML 与 compose 覆盖，标记为 `Configured`，"
-        "不代表训练验证。"
         "mjwarp playback 默认仅支持显式、有限步数的 `record` 并复用 MuJoCo 离线 renderer；"
         "`uv run eval --sim mjwarp --render-mode interactive` 路由到 MuJoCo 交互 viewer"
         "（mjwarp 跑物理、MuJoCo 渲染 env[0]，强制单 env）；不支持 `auto` 或 native playback。"

src/unilab/assets/hub.py

@@ -54,10 +54,6 @@
     "go2w": (("robots/go2/assets", "base_0.obj", "**/*", "asset"),),
     "microduck": (("robots/microduck/assets", "trunk_base.stl", "*.stl", "STL"),),
     "sharpa_wave": (("robots/sharpa_wave/meshes", "DP_HB1_4F.STL", "*.STL", "STL"),),
-    "t800": (
-        ("robots/t800/assets", "LINK_BASE.obj", "*.obj", "OBJ"),
-        ("robots/t800/textures", "LINK_BASE.png", "*.png", "PNG"),
-    ),
     "x2": (("robots/x2/meshes", "pelvis.STL", "*.STL", "STL"),),
 }