feat!: move G1 walk rough downstream

docs/sphinx/source/en/2-user_guide/2-algorithms/3-sac.md

@@ -20,7 +20,6 @@ CPU and XPU training are unsupported; there is no alternate replay pipeline.
 
 ```bash
 uv run train --algo sac --task g1_walk_flat --sim mujoco
-uv run train --algo sac --task g1_walk_rough --sim motrix training.no_play=true
 ```
 
 ## Key Fields

docs/sphinx/source/en/2-user_guide/4-tasks/1-locomotion.md

@@ -7,7 +7,7 @@ Manager-Based runtime. Unitree production variants live in the downstream
 ## Core reference tasks
 
 - Go2 joystick: `go2_joystick_flat`
-- G1 walking: `g1_walk_flat`, `g1_walk_rough`
+- G1 walking: `g1_walk_flat`
 - G1 motion tracking reference profiles: see {doc}`2-motion_tracking`
 
 ```bash

docs/sphinx/source/en/2-user_guide/5-domain_randomization/0-index.md

@@ -33,7 +33,6 @@ These three paths correspond to three lifecycle classes:
 | --- | --- | --- | --- | --- | --- |
 | `Go2JoystickFlat` | Hydra `events:` terms | Yes: owner YAML declares reset events | root-state reset + `pd_gains` kp/kd | none | `src/unilab/conf/ppo/task/go2_joystick_flat/base.yaml` |
 | `G1WalkFlat` | Hydra `events:` terms | Yes: Hydra `EventTermCfg` + Manager-Based reset terms | root-state reset + kp/kd via `pd_gains` | none | `g1/manager_terms.py` |
-| `G1WalkRough` | Hydra `events:` terms | Yes: same Manager-Based event terms as `G1WalkFlat` | root-state reset + kp/kd via `pd_gains` | none | `g1/manager_terms.py` |
 | `G1MotionTracking` | Hydra command term | Yes: Hydra `MotionCommandCfg` + Manager-Based command reset | motion frame, root pose/velocity, and joint-position sampling | none | `motion_tracking/common/manager_terms.py` |
 | `G1WBTObs` | Hydra `events:` terms | Yes: same motion command + Hydra `EventTermCfg` | motion reset plus mass/COM/PD/friction/encoder-bias events | interval velocity kick | `motion_tracking/g1/manager_terms.py` |
 | `AllegroInhandRotation` | Hydra `events:` terms | Yes: Hydra `EventTermCfg` + Manager-Based reset term | entity-scoped hand/ball reset | none | `allegro_inhand/manager_terms.py` |
@@ -45,7 +44,6 @@ These three paths correspond to three lifecycle classes:
 | --- | --- | --- | --- |
 | `Go2JoystickFlat` | base xy/yaw and base qvel via `reset_root_state_uniform`; command sampling; kp/kd via `pd_gains` | none | event terms declared and enabled by default in `src/unilab/conf/ppo/task/go2_joystick_flat/base.yaml` |
 | `G1WalkFlat` | base xy/yaw and base qvel via `reset_root_state_uniform`; command sampling with a planar dead zone; `gait_phase` sampling; kp/kd randomization via `pd_gains` | none | kp/kd enabled on mujoco owners by default; disabled on motrix/mjwarp owners |
-| `G1WalkRough` | Same as `G1WalkFlat` (shared owner bases, rough scene) | none | Same defaults as `G1WalkFlat` |
 | `G1MotionTracking` | Motion-command frame sampling; root pose perturbation `x/y/z/roll/pitch/yaw`; root velocity perturbation `x/y/z/roll/pitch/yaw`; joint-position noise clipped through the public entity soft limits; action-manager state reset | none | `pose_range`, `velocity_range`, and `joint_position_range` have non-zero perturbations in the base owner |
 | `G1WBTObs` | Same motion reset plus base mass, base COM, PD gain, foot friction, and encoder-bias event terms | `push_by_setting_velocity` | The WBT owner explicitly enables all listed event terms; unsupported capabilities raise rather than fall back |
 | `AllegroInhandRotation` | Entity-scoped hand/ball reset; an explicitly configured grasp cache is sampled, otherwise `null` explicitly selects the model home pose; optional `joint_noise`, `ball_velocity_noise`, and `ball_z_offset` | none | owner YAML explicitly selects the home pose and zero reset noise; a configured missing or malformed cache fails closed |
@@ -132,7 +130,6 @@ The ownership boundary and MJWarp/CPU executor split are recorded in
 ## Related Tasks
 
 - {doc}`G1 Motion Tracking <../4-tasks/2-motion_tracking>`: confirm motion assets and replay first before enabling DR.
-- {doc}`Go2 Rough Terrain <../4-tasks/1-locomotion>`: common items are mass, COM, friction, and push.
 
 For the backend capability boundary, see
 {doc}`Domain Randomization Contract </en/4-developer_guide/2-contracts/4-dr_contract>`.

docs/sphinx/source/en/2-user_guide/6-terrain/1-procedural.md

@@ -28,10 +28,10 @@ env:
         seed: 42
 ```
 
-UniLab does not retain a procedural-terrain production owner. The retained
-`g1_walk_rough` task uses the static-heightfield form of the same cold-path
-scene contract, while generator composition itself is covered by core terrain
-and backend materialization tests.
+UniLab does not retain a procedural-terrain production owner. Generator
+composition and backend materialization are covered by core terrain and
+materialization tests; task owners that enable the capability carry their own
+training evidence.
 
 ## Materialization boundary
 
@@ -47,15 +47,10 @@ During `registry.make(...)`:
 Step and reset never parse robot XML or inspect asset metadata. They consume
 cached IDs and public backend capabilities.
 
-## Example
+## Owner integration
 
-Use the retained rough-scene reference task:
-
-```bash
-uv run train --algo sac --task g1_walk_rough --sim mujoco training.no_play=true
-```
-
-Custom owners can tune the terrain grid and seed through Hydra:
+Select a task owner that declares `env.scene.terrain`, then tune the terrain
+grid and seed through Hydra:
 
 ```text
 env.scene.terrain.generator.num_rows=4

docs/sphinx/source/en/2-user_guide/6-terrain/2-heightfield_import.md

@@ -1,25 +1,23 @@
 # Heightfield Import
 
 Heightfield terrain is configured through `SceneCfg` and the terrain generator,
-then materialized by the backend on the init path. The committed user-facing
-example is the retained `g1_walk_rough` scene.
+then materialized by the backend on the init path. Core tests exercise the
+contract with a generated heightfield scene.
 
 ## Files To Read
 
 - `src/unilab/terrains/heightfield_terrains.py`
 - `unisim.terrain.generator`
-- `src/unilab/assets/robots/g1/scene_rough.xml`
 - `src/unilab/tasks/locomotion/common/height_scan.py`
+- `tests/utils/test_xml_utils.py`
 - `unisim.backend.mujoco.xml`
 - `unisim.backend.motrix.scene`
 
 ## Smoke Commands
 
 ```bash
-uv run train --algo sac --task g1_walk_rough --sim mujoco \
-  algo.max_iterations=2 \
-  algo.num_envs=64 \
-  training.no_play=true
+uv run pytest tests/utils/test_xml_utils.py \
+  -k materialize_mujoco_hfield_attached_scene
 ```
 
 Height scan IDs and offsets are cached during env initialization; hot paths call

docs/sphinx/source/en/4-developer_guide/1-architecture/4-scene_composition.md

@@ -101,14 +101,14 @@ Disallowed on hot paths:
 - Probing backend-private scene methods instead of using explicit contracts.
 - Regenerating terrain after env construction.
 
-## Go2 Rough Terrain Evidence
+## Terrain Contract Evidence
 
-The current procedural terrain user-facing path is Go2 rough terrain:
+The core repository keeps the terrain contract and its materialization tests,
+but it does not own a production rough-terrain task:
 
-- Task owner: `src/unilab/tasks/locomotion/g1/manager_terms.py`
 - Terrain generator: `unisim.terrain.generator`
 - MuJoCo materializer: `unisim.backend.mujoco.xml`
 - Motrix materializer: `unisim.backend.motrix.scene`
-- Owner YAMLs: `src/unilab/conf/sac/task/g1_walk_rough/{mujoco,motrix}.yaml`
+- Contract coverage: `tests/utils/test_xml_utils.py`
 
 User instructions are in {doc}`../../2-user_guide/6-terrain/1-procedural`.

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

@@ -94,7 +94,6 @@ rendering/playback paths remain unsupported.
 | APPO (torch) | `g1_flip_tracking` (G1 flip tracking) | Tested | - | Tested | - | - | - | - | - |
 | APPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Tested | Tested | Tested | Tested | Configured | Tested | - |
-| SAC (torch) | `g1_walk_rough` (G1 walk rough) | Tested | - | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Configured | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_flip_tracking` (G1 flip tracking) | Tested | - | Registered | - | - | - | - | - |
 | SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | Registered | - | - | - | - | - |

docs/sphinx/source/zh_CN/2-user_guide/2-algorithms/3-sac.md

@@ -17,7 +17,6 @@ submission。CPU 与 XPU training 不受支持，也不存在第二套 replay pi
 
 ```bash
 uv run train --algo sac --task g1_walk_flat --sim mujoco
-uv run train --algo sac --task g1_walk_rough --sim motrix training.no_play=true
 ```
 
 ## 关键字段

docs/sphinx/source/zh_CN/2-user_guide/4-tasks/1-locomotion.md

@@ -6,7 +6,7 @@ runtime。Unitree production 变体由下游 `unitree_rl_unilab` 包维护。
 ## 核心参考任务
 
 - Go2 joystick: `go2_joystick_flat`
-- G1 walking: `g1_walk_flat`, `g1_walk_rough`
+- G1 walking: `g1_walk_flat`
 - G1 motion tracking 参考配置：见 {doc}`2-motion_tracking`
 
 ```bash

docs/sphinx/source/zh_CN/2-user_guide/5-domain_randomization/0-index.md

@@ -29,7 +29,6 @@ Manager-Based event term 是唯一 DR 声明路径：
 | --- | --- | --- | --- | --- | --- |
 | `Go2JoystickFlat` | Hydra `events:` term | 是：owner YAML 声明 reset event | root-state reset + `pd_gains` kp/kd | 无 | `src/unilab/conf/ppo/task/go2_joystick_flat/base.yaml` |
 | `G1WalkFlat` | Hydra `events:` term | 是：Hydra `EventTermCfg` + Manager-Based reset term | root-state reset + 经 `pd_gains` 的 kp/kd | 无 | `g1/manager_terms.py` |
-| `G1WalkRough` | Hydra `events:` term | 是：与 `G1WalkFlat` 相同的 Manager-Based event term | root-state reset + 经 `pd_gains` 的 kp/kd | 无 | `g1/manager_terms.py` |
 | `G1MotionTracking` | Hydra command term | 是：Hydra `MotionCommandCfg` + Manager-Based command reset | motion frame、root pose/velocity 与 joint-position 采样 | 无 | `motion_tracking/common/manager_terms.py` |
 | `G1WBTObs` | Hydra `events:` term | 是：同一 motion command + Hydra `EventTermCfg` | motion reset 加 mass/COM/PD/friction/encoder-bias event | interval velocity kick | `motion_tracking/g1/manager_terms.py` |
 | `AllegroInhandRotation` | Hydra `events:` term | 是：Hydra `EventTermCfg` + Manager-Based reset term | entity 范围的手/球 reset | 无 | `allegro_inhand/manager_terms.py` |
@@ -41,7 +40,6 @@ Manager-Based event term 是唯一 DR 声明路径：
 | --- | --- | --- | --- |
 | `Go2JoystickFlat` | 经 `reset_root_state_uniform` 的 base xy/yaw 与 base qvel；command 采样；经 `pd_gains` 的 kp/kd | 无 | event term 在 `src/unilab/conf/ppo/task/go2_joystick_flat/base.yaml` 中默认声明并启用 |
 | `G1WalkFlat` | 经 `reset_root_state_uniform` 的 base xy/yaw 与 base qvel；带平面死区的 command 采样；`gait_phase` 采样；经 `pd_gains` 的 kp/kd 随机化 | 无 | mujoco owner 默认启用 kp/kd；motrix/mjwarp owner 默认禁用 |
-| `G1WalkRough` | 与 `G1WalkFlat` 相同（共享 owner base，rough 场景） | 无 | 与 `G1WalkFlat` 相同的默认值 |
 | `G1MotionTracking` | Motion-command frame 采样；root 位姿扰动 `x/y/z/roll/pitch/yaw`；root 速度扰动 `x/y/z/roll/pitch/yaw`；通过 public entity soft limit clip 的关节位置噪声；action-manager 状态 reset | 无 | base owner 中 `pose_range`、`velocity_range` 与 `joint_position_range` 默认有非零扰动 |
 | `G1WBTObs` | 同一 motion reset 加 base mass、base COM、PD gain、足端摩擦和 encoder-bias event term | `push_by_setting_velocity` | WBT owner 显式启用上述全部 event term；能力不支持时直接报错，不回退 |
 | `AllegroInhandRotation` | entity 范围的手/球 reset；显式配置 grasp cache 时进行采样，否则以 `null` 显式选择模型 home pose；可选 `joint_noise`、`ball_velocity_noise` 与 `ball_z_offset` | 无 | owner YAML 显式选择 home pose 与零 reset 噪声；配置的 cache 缺失或格式错误时 fail-closed |
@@ -119,7 +117,6 @@ Reset-time model-field DR 保持在已选 identity 内。其 canonical 或 per-e
 ## 相关任务
 
 - {doc}`G1 Motion Tracking <../4-tasks/2-motion_tracking>`：开启 DR 前先确认 motion 资产和 replay。
-- {doc}`Go2 Rough Terrain <../4-tasks/1-locomotion>`：常见的是 mass、COM、friction、push。
 
 有关后端能力边界，请参阅
 {doc}`Domain Randomization Contract </zh_CN/4-developer_guide/2-contracts/4-dr_contract>`。

docs/sphinx/source/zh_CN/2-user_guide/6-terrain/1-procedural.md

@@ -26,9 +26,9 @@ env:
         seed: 42
 ```
 
-UniLab 不保留 procedural-terrain production owner。保留的 `g1_walk_rough`
-任务使用同一冷路径 scene contract 的静态 heightfield 形式；generator 组合
-本身由核心 terrain 与 backend materialization 测试覆盖。
+UniLab 不保留 procedural-terrain production owner。generator 组合与 backend
+materialization 由核心 terrain 和 materialization 测试覆盖；启用该能力的
+task owner 需要自行持有训练证据。
 
 ## 物化边界
 
@@ -44,15 +44,10 @@ UniLab 不保留 procedural-terrain production owner。保留的 `g1_walk_rough`
 step 和 reset 不解析机器人 XML，也不检查 asset metadata；它们只消费缓存
 ID 和公开 backend capability。
 
-## 示例
+## Owner 集成
 
-使用保留的 rough-scene reference task：
-
-```bash
-uv run train --algo sac --task g1_walk_rough --sim mujoco training.no_play=true
-```
-
-自定义 owner 可以通过 Hydra 调整 terrain 网格与 seed：
+选择声明了 `env.scene.terrain` 的 task owner，然后通过 Hydra 调整 terrain
+网格与 seed：
 
 ```text
 env.scene.terrain.generator.num_rows=4

docs/sphinx/source/zh_CN/2-user_guide/6-terrain/2-heightfield_import.md

@@ -1,23 +1,21 @@
 # 高度场导入
 
-高度场地形通过 `SceneCfg` 和地形生成器进行配置，然后在 init 路径上由后端实例化。已提交的面向用户示例是保留的 `g1_walk_rough` 场景。
+高度场地形通过 `SceneCfg` 和地形生成器进行配置，然后在 init 路径上由后端实例化。核心测试使用生成式 heightfield 场景覆盖该契约。
 
 ## 需要阅读的文件
 
 - `src/unilab/terrains/heightfield_terrains.py`
 - `unisim.terrain.generator`
-- `src/unilab/assets/robots/g1/scene_rough.xml`
 - `src/unilab/tasks/locomotion/common/height_scan.py`
+- `tests/utils/test_xml_utils.py`
 - `unisim.backend.mujoco.xml`
 - `unisim.backend.motrix.scene`
 
 ## 冒烟命令
 
 ```bash
-uv run train --algo sac --task g1_walk_rough --sim mujoco \
-  algo.max_iterations=2 \
-  algo.num_envs=64 \
-  training.no_play=true
+uv run pytest tests/utils/test_xml_utils.py \
+  -k materialize_mujoco_hfield_attached_scene
 ```
 
 高度扫描的 ID 和偏移在 env 初始化期间缓存；热路径调用后端高度 scanner contract，而不是解析 XML 或 asset 元数据。

docs/sphinx/source/zh_CN/4-developer_guide/1-architecture/4-scene_composition.md

@@ -94,14 +94,14 @@ materializer。
 - 探测 backend 私有的场景方法，而不使用明确的契约。
 - 在 env 构造完成后重新生成地形。
 
-## Go2 崎岖地形证据
+## 地形契约证据
 
-当前面向用户的程序化地形路径是 Go2 崎岖地形：
+核心仓库保留 terrain contract 与 materialization 测试，但不再持有
+production rough-terrain 任务：
 
-- Task owner：`src/unilab/tasks/locomotion/g1/manager_terms.py`
 - 地形生成器：`unisim.terrain.generator`
 - MuJoCo materializer：`unisim.backend.mujoco.xml`
 - Motrix materializer：`unisim.backend.motrix.scene`
-- Owner YAML：`src/unilab/conf/sac/task/g1_walk_rough/{mujoco,motrix}.yaml`
+- 契约覆盖：`tests/utils/test_xml_utils.py`
 
 用户使用说明见 {doc}`../../2-user_guide/6-terrain/1-procedural`。

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

@@ -77,7 +77,6 @@ uv run scripts/generate_support_matrix.py --write
 | APPO (torch) | `g1_flip_tracking` (G1 flip tracking) | Tested | - | Tested | - | - | - | - | - |
 | APPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Tested | Tested | Tested | Tested | Configured | Tested | - |
-| SAC (torch) | `g1_walk_rough` (G1 walk rough) | Tested | - | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Configured | Tested | - | - | - | - | - |
 | SAC (torch) | `g1_flip_tracking` (G1 flip tracking) | Tested | - | Registered | - | - | - | - | - |
 | SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | Registered | - | - | - | - | - |

scripts/benchmark/env/benchmark_env_step.py

@@ -6,7 +6,6 @@
 
     # Single task + backend:
     uv run scripts/benchmark/env/benchmark_env_step.py task=g1_walk_flat/motrix
-    uv run scripts/benchmark/env/benchmark_env_step.py task=g1_walk_rough/mujoco
 
     # mjwarp backend (Phase 1: g1_walk_flat only) requires extra deps:
     uv run --with mujoco-warp --with warp-lang \\
@@ -185,44 +184,6 @@ def _ppo_owner_yaml_cfg(
     )
 
 
-def _sac_owner_yaml_cfg(
-    task_id: str,
-    backend: str,
-    env_cfg_cls: Callable[[], Any],
-    config_overrides: list[str],
-) -> Any:
-    yaml_backend = _hydra_yaml_backend(backend)
-    return _owner_yaml_cfg(
-        config_root="sac",
-        algo_name="sac",
-        overrides=[f"task={task_id}/{yaml_backend}"],
-        config_overrides=config_overrides,
-        env_cfg_cls=env_cfg_cls,
-    )
-
-
-def _materialize_g1_rough_benchmark_scene() -> str:
-    import shutil
-    import xml.etree.ElementTree as ET
-
-    source_dir = ROOT_DIR / "src" / "unilab" / "assets" / "robots" / "g1"
-    output_dir = Path("/tmp/unilab_benchmark_g1_rough_scene")
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    for name in ("g1.xml",):
-        shutil.copy2(source_dir / name, output_dir / name)
-    for name in ("assets", "textures", "hfields"):
-        shutil.copytree(source_dir / name, output_dir / name, dirs_exist_ok=True)
-
-    tree = ET.parse(source_dir / "scene_rough.xml")
-    hfield = tree.getroot().find("./asset/hfield[@name='hfield']")
-    if hfield is not None:
-        hfield.set("file", "hfields/hfield.png")
-    output_path = output_dir / "scene_rough.xml"
-    tree.write(output_path)
-    return str(output_path)
-
-
 def _manager_env_cls() -> Callable[..., Any]:
     from unilab.envs import make_manager_based_rl_env
 
@@ -241,12 +202,6 @@ def _g1_flat_cfg(backend: str, config_overrides: list[str]) -> Any:
     return _ppo_owner_yaml_cfg("g1_walk_flat", backend, ManagerBasedRlEnvCfg, config_overrides)
 
 
-def _g1_rough_cfg(backend: str, config_overrides: list[str]) -> Any:
-    from unilab.envs import ManagerBasedRlEnvCfg
-
-    return _sac_owner_yaml_cfg("g1_walk_rough", backend, ManagerBasedRlEnvCfg, config_overrides)
-
-
 def _g1_motion_tracking_cfg(backend: str, config_overrides: list[str]) -> Any:
     from unilab.envs import ManagerBasedRlEnvCfg
 
@@ -279,14 +234,6 @@ def _g1_walk_env_cls() -> type:
         env_cls_factory=_g1_walk_env_cls,
         backends=("mujoco", "motrix", "mjwarp"),
     ),
-    "g1_rough": TaskConfig(
-        task_id="g1_walk_rough",
-        env_name="G1WalkRough",
-        cfg_factory=_g1_rough_cfg,
-        env_cls_factory=_g1_walk_env_cls,
-        aliases=("sac/g1_walk_rough",),
-        backends=("mujoco", "motrix", "mjwarp"),
-    ),
     "g1_mt": TaskConfig(
         task_id="g1_motion_tracking",
         env_name="G1MotionTracking",
@@ -305,7 +252,6 @@ def _g1_walk_env_cls() -> type:
     "go2": "#54A24B",
     "g1": "#F58518",
     "g1_mt": "#B279A2",
-    "g1_rough": "#E45756",
 }
 BACKEND_STYLES = {
     "mujoco": {"marker": "o", "linestyle": "-", "hatch": "//"},
@@ -653,8 +599,6 @@ def _short_task_label(task_name: str) -> str:
     name = task_name.lower()
     if "motiontracking" in name:
         return "g1_mt"
-    if "rough" in name and name.startswith("g1"):
-        return "g1_rough"
     for prefix in ("go2", "g1"):
         if name.startswith(prefix):
             return prefix

scripts/tools/support_matrix.py

@@ -81,17 +81,15 @@
 _TASK_ORDER = {
     "go2_joystick_flat": 0,
     "g1_walk_flat": 1,
-    "g1_walk_rough": 2,
-    "g1_motion_tracking": 3,
-    "g1_flip_tracking": 4,
-    "x2_wall_flip_tracking": 5,
-    "allegro_inhand": 6,
-    "allegro_sac": 7,
+    "g1_motion_tracking": 2,
+    "g1_flip_tracking": 3,
+    "x2_wall_flip_tracking": 4,
+    "allegro_inhand": 5,
+    "allegro_sac": 6,
 }
 _TASK_LABELS = {
     "go2_joystick_flat": "Go2 joystick",
     "g1_walk_flat": "G1 walk flat",
-    "g1_walk_rough": "G1 walk rough",
     "g1_motion_tracking": "G1 motion tracking",
     "g1_flip_tracking": "G1 flip tracking",
     "x2_wall_flip_tracking": "X2 wall flip tracking",