feat(appo): restore Go2 Motrix owner on tensor runtime

docs/sphinx/source/en/2-user_guide/3-backends/2-motrix.md

@@ -5,17 +5,18 @@ Motrix is an optional backend installed through the `motrix` extra, which
 delegates to `unisim-core[motrix]`: the runtime pin lives in UniSim's
 `pyproject.toml`, and the adapter lives under `unisim.backend.motrix`.
 
-Motrix is a CPU-authoritative packed HOST_BRIDGE backend. Issue #2054 restores
-it to the tensor-only Manager runtime for exactly two canonical training
-workloads:
+Motrix is a CPU-authoritative packed HOST_BRIDGE backend. Issue #2054 restored
+it to the tensor-only Manager runtime with these canonical training workloads:
 
 ```bash
+uv run --extra motrix train --algo appo --task go2_joystick_flat --sim motrix
 uv run --extra motrix train --algo sac --task g1_walk_flat --sim motrix
 uv run --extra motrix train --algo flashsac --task g1_motion_tracking --sim motrix
 ```
 
 Other Motrix owners remain out of scope and do not constitute production
-support claims.
+support claims. The Go2 APPO owner is restored as a configured training path;
+its current evidence is bounded smoke training, not a performance benchmark.
 
 The public tensor lifecycle exposes:
 
@@ -38,7 +39,7 @@ uv sync --extra motrix
 
 ## When To Use It
 
-- The workload is one of the two canonical owners above.
+- The workload is one of the canonical owners above.
 - You need a CPU-authoritative packed HOST_BRIDGE comparison/reference path.
 - The generated support matrix marks the entrypoint/task/backend combination as
   configured or tested.

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

@@ -86,15 +86,15 @@ This table is derived from UniSim's SDK-free public static inventory. It describ
 
 | Entrypoint | Task owner | MuJoCo | Motrix | Drake | mjwarp | Newton | SuperDex | Genesis |
 |------------|------------|---|---|---|---|---|---|---|
-| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Configured | - | - | Configured | - |
+| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Registered | Configured | - | - | Configured | - |
 | PPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Tested | Registered | - | Configured |
 | PPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Registered | Registered | - | Registered |
-| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Registered | - | - | Registered | - |
+| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Tested | Registered | - | - | Registered | - |
 | APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Registered | Registered | - | Registered |
 | APPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Registered | Registered | - | Registered |
 | SAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Tested | - | Tested | Tested | - | Tested |
 | SAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Configured | Registered | - | Configured |
-| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Registered | - | - | Registered | - |
+| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Registered | Registered | - | - | Registered | - |
 | FlashSAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Configured | Registered | - | Registered |
 | FlashSAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Tested | - | Configured | Configured | - | Configured |
 

docs/sphinx/source/zh_CN/2-user_guide/3-backends/2-motrix.md

@@ -5,14 +5,17 @@ Motrix 是一个可选后端，通过 `motrix` extra 安装。该 extra 委托
 中，适配层位于 `unisim.backend.motrix` 下。
 
 Motrix 是 CPU-authoritative packed HOST_BRIDGE 后端。issue #2054 将其恢复
-进 tensor-only Manager runtime，但范围仅限两个 canonical workload：
+进 tensor-only Manager runtime，当前包含这些 canonical workload：
 
 ```bash
+uv run --extra motrix train --algo appo --task go2_joystick_flat --sim motrix
 uv run --extra motrix train --algo sac --task g1_walk_flat --sim motrix
 uv run --extra motrix train --algo flashsac --task g1_motion_tracking --sim motrix
 ```
 
 其他 Motrix owner 仍不在支持范围内，也不构成生产支持声明。
+Go2 APPO owner 是已恢复的 configured 训练路径；当前证据仅限 bounded smoke
+training，不构成性能 benchmark。
 
 public tensor lifecycle 提供：
 
@@ -35,7 +38,7 @@ uv sync --extra motrix
 
 ## 何时使用
 
-- workload 是上述两个 canonical owner 之一。
+- workload 是上述 canonical owner 之一。
 - 需要 CPU-authoritative packed HOST_BRIDGE 对比/参考路径。
 - 所生成的支持矩阵将你的 entrypoint/task/backend 组合标记为 configured 或 tested。
 

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

@@ -77,15 +77,15 @@ uv run scripts/generate_support_matrix.py --write
 
 | Entrypoint | Task owner | MuJoCo | Motrix | Drake | mjwarp | Newton | SuperDex | Genesis |
 |------------|------------|---|---|---|---|---|---|---|
-| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Configured | - | - | Configured | - |
+| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Registered | Configured | - | - | Configured | - |
 | PPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Tested | Registered | - | Configured |
 | PPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Registered | Registered | - | Registered |
-| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Registered | - | - | Registered | - |
+| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Tested | Registered | - | - | Registered | - |
 | APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Registered | Registered | - | Registered |
 | APPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Registered | Registered | - | Registered |
 | SAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Tested | - | Tested | Tested | - | Tested |
 | SAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Configured | Registered | - | Configured |
-| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Registered | - | - | Registered | - |
+| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | Registered | Registered | - | - | Registered | - |
 | FlashSAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Configured | Registered | - | Registered |
 | FlashSAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Tested | - | Configured | Configured | - | Configured |
 

pyproject.toml

@@ -107,7 +107,7 @@ newton = [
 ]
 motrix = ["unisim-core[motrix]>=1.7.11"]
 genesis = ["genesis-world==1.3.3"]
-uni_rl = ["unilab-rl==1.4.6"]
+uni_rl = ["unilab-rl==1.4.8"]
 # Wheels are CPython 3.12/3.13 Linux x86_64 only; the marker keeps uv lock
 # resolvable for the other required-environments.
 superdex = [
@@ -117,7 +117,7 @@ superdex = [
 [dependency-groups]
 # Dev/CI exercises the APPO/off-policy/multi-GPU paths, so the optional
 # unilab-rl runtime stays installed in the dev environment.
-dev = ["pytest", "pytest-cov", "ruff", "mypy", "pyright>=1.1.408", "unilab-rl==1.4.6"]
+dev = ["pytest", "pytest-cov", "ruff", "mypy", "pyright>=1.1.408", "unilab-rl==1.4.8"]
 
 [[tool.uv.index]]
 name = "r2-cu130"

src/unilab/conf/appo/task/go2_joystick_flat/motrix.yaml

@@ -0,0 +1,21 @@
+# @package _global_
+defaults:
+  - /task/go2_joystick_flat/base
+  - _self_
+
+training:
+  task_name: Go2JoystickFlat
+  sim_backend: motrix
+algo:
+  num_envs: 512
+  steps_per_env: 24
+  max_iterations: 180
+env:
+  # Motrix's packed HOST_BRIDGE uses the canonical tensor selected-reset terms.
+  events:
+    reset_scene_to_default:
+      func: unilab.envs.mdp.reset_scene_to_default_tensor
+    reset_root_state_uniform:
+      func: unilab.envs.mdp.reset_root_state_uniform_tensor
+    # Motrix tensor selected reset does not support PD-gain randomization.
+    pd_gains: null

src/unilab/scripts/train_appo.py

@@ -108,6 +108,9 @@ def build_appo_runner_kwargs(
     nan_guard_cfg = resolve_nan_guard_cfg(cfg.training)
     if nan_guard_cfg is not None:
         runner_kwargs["nan_guard_cfg"] = nan_guard_cfg
+        from unilab.training.tensor_runtime_hooks import build_tensor_nan_guard_factory
+
+        runner_kwargs["nan_guard_factory"] = build_tensor_nan_guard_factory()
     return runner_kwargs
 
 

src/unilab/tasks/locomotion/go2/__init__.py

@@ -5,6 +5,7 @@
 
 registry.register_env_config("Go2JoystickFlat", ManagerBasedRlEnvCfg)
 registry.register_env("Go2JoystickFlat", make_manager_based_rl_env, sim_backend="mujoco")
+registry.register_env("Go2JoystickFlat", make_manager_based_rl_env, sim_backend="motrix")
 registry.register_env("Go2JoystickFlat", make_manager_based_rl_env, sim_backend="superdex")
 registry.register_env("Go2JoystickFlat", make_manager_based_rl_env, sim_backend="drake")
 

tests/algos/test_appo_runner.py

@@ -53,6 +53,7 @@ def test_appo_runner_learn_two_iterations(env_name):
         env_factory=registry_env_factory(env_name, "mujoco"),
         env_cfg_overrides=env_cfg_overrides,
         rl_cfg=cfg,
+        device="cpu",
         num_envs=128,
         steps_per_env=8,
     )

tests/envs/locomotion/go2/test_manager_based_cfg.py

@@ -76,6 +76,16 @@
         False,
         id="appo-mujoco",
     ),
+    pytest.param(
+        "appo",
+        ("task=go2_joystick_flat/motrix",),
+        "motrix",
+        0.25,
+        False,
+        True,
+        False,
+        id="appo-motrix",
+    ),
     pytest.param(
         "flashsac",
         ("task=go2_joystick_flat/mujoco",),
@@ -228,7 +238,7 @@ def test_go2_flat_registry_is_manager_only() -> None:
     assert bare_cfg.rewards == {}
     assert registry.list_registered_envs()["Go2JoystickFlat"] == {
         "config_factory": "ManagerBasedRlEnvCfg",
-        "available_backends": ["mujoco", "superdex", "drake"],
+        "available_backends": ["mujoco", "motrix", "superdex", "drake"],
     }
 
 
@@ -255,6 +265,15 @@ def test_go2_flat_drake_owner_keeps_policy_io_and_disables_unsupported_dr() -> N
     ]
 
 
+def test_go2_flat_motrix_owner_uses_tensor_reset_events() -> None:
+    registry.ensure_registries()
+    _, env_cfg, _ = _materialize("appo", ("task=go2_joystick_flat/motrix",))
+
+    assert env_cfg.events["reset_scene_to_default"].func is mdp.reset_scene_to_default_tensor
+    assert env_cfg.events["reset_root_state_uniform"].func is mdp.reset_root_state_uniform_tensor
+    assert env_cfg.events["pd_gains"] is None
+
+
 @pytest.mark.parametrize(
     ("backend", "owner"),
     (("mujoco", "task=go2_joystick_flat/mujoco"),),

tests/scripts/test_support_matrix.py

@@ -86,7 +86,9 @@ def test_support_matrix_marks_go2_ppo_backends_as_tested():
 
     assert row.cells["mujoco"].level == EvidenceLevel.TESTED
     assert row.cells["mjwarp"].level == EvidenceLevel.MISSING
-    assert row.cells["motrix"].level is EvidenceLevel.MISSING
+    assert row.cells["motrix"].level is EvidenceLevel.REGISTERED
+    appo_row = _row("APPO (torch)", "go2_joystick_flat")
+    assert appo_row.cells["motrix"].level is EvidenceLevel.TESTED
 
 
 def test_support_matrix_marks_validated_g1_mjwarp_entrypoints_as_tested():

tests/scripts/test_train_script_configs.py

@@ -29,6 +29,10 @@
     "g1_motion_tracking/mujoco",
 ]
 
+APPO_MOTRIX_SMOKE_TASKS = [
+    "go2_joystick_flat/motrix",
+]
+
 APPO_MOTION_SMOKE_TASKS = {
     "g1_motion_tracking/mujoco",
 }
@@ -97,6 +101,32 @@ def test_appo_task_configs_load(task, tmp_path):
     assert result.returncode == 0, f"APPO {task} failed:\n{result.stderr}"
 
 
+@pytest.mark.slow
+@pytest.mark.parametrize("task", APPO_MOTRIX_SMOKE_TASKS)
+def test_appo_motrix_task_configs_load(task, tmp_path):
+    """APPO can complete a bounded Motrix training smoke."""
+    pytest.importorskip("motrixsim", reason="motrix extra not available")
+    result = subprocess.run(
+        [
+            sys.executable,
+            "src/unilab/scripts/train_appo.py",
+            f"task={task}",
+            "algo.num_envs=16",
+            "algo.steps_per_env=4",
+            "algo.max_iterations=1",
+            "algo.save_interval=100",
+            "training.no_play=true",
+            f"training.log_root={tmp_path}",
+        ],
+        capture_output=True,
+        text=True,
+        timeout=180,
+    )
+    assert result.returncode == 0, f"APPO {task} failed:\n{result.stderr}"
+    checkpoints = list(tmp_path.glob("Go2JoystickFlat/*/model_1.pt"))
+    assert len(checkpoints) == 1, checkpoints
+
+
 @pytest.mark.slow
 @pytest.mark.parametrize(
     ("algo", "task"),

tests/tasks/test_go2_tensor_runtime.py

@@ -352,6 +352,7 @@ def test_go2_runtime_does_not_add_backend_registrations() -> None:
     registry.ensure_registries()
     assert registry.list_registered_envs()["Go2JoystickFlat"]["available_backends"] == [
         "mujoco",
+        "motrix",
         "superdex",
         "drake",
     ]