feat(drake): scope Go2 owner into tensor runtime

Makefile

@@ -89,3 +89,7 @@ clean:
 	find src/unilab/assets/caches -type f ! -name '.gitkeep' -delete 2>/dev/null || true
 	find src/unilab/assets/checkpoints -type f ! -name '.gitkeep' -delete 2>/dev/null || true
 	find src/unilab/assets/scenes -type f ! -name '.gitkeep' -delete 2>/dev/null || true
+
+.PHONY: setup-drake
+setup-drake:
+	uv run --no-sync bash scripts/tools/setup_drake_env.sh --download-drake

docs/sphinx/source/en/1-getting_started/2-installation.md

@@ -120,9 +120,9 @@ is unreachable.
 
 ## Backend Extras
 
-The tensor-only Manager runtime currently exposes five production backends:
-`mujoco`, `mjwarp`, `genesis`, `newton`, and `motrix`. Their simulator
-dependencies are optional.
+The tensor-only Manager runtime currently exposes five production backends
+(`mujoco`, `mjwarp`, `genesis`, `newton`, and `motrix`) plus the scoped Drake
+owner. Their simulator dependencies are optional.
 
 ```bash
 # All scoped backends together.
@@ -136,11 +136,12 @@ uv sync --extra mujoco --extra mjwarp --extra genesis --extra newton --extra mot
 | Genesis | `uv sync --extra genesis` | The validated path uses Linux x86_64, an NVIDIA GPU, and the pinned torch/Genesis versions |
 | Newton | `uv sync --extra newton` | Linux CUDA with the pinned newton / MuJoCo-Warp / Warp versions |
 | Motrix | `uv sync --extra motrix` | CPU-authoritative MotrixSim with packed Torch HOST_BRIDGE transfers |
+| Drake | `make setup-drake` | Local Drake C++ prefix plus the DrakeUni native batch extension; currently scoped to PPO `go2_joystick_flat` |
 
-The `drake`, `isaacgym`, `isaacsim`, and `superdex` adapters remain temporarily
-shelved by `unisim-core` during issue #1811. Their extras and historical backend
-pages are not production support claims and the UniLab train/eval CLI rejects
-them until new capability, parity, and support-matrix evidence is provided.
+The `isaacgym`, `isaacsim`, and `superdex` adapters remain temporarily shelved
+by `unisim-core` during issue #1811. Their extras and historical backend pages
+are not production support claims and the UniLab train/eval CLI rejects them
+until new capability, parity, and support-matrix evidence is provided.
 
 Read the scoped backend pages for runtime variables, renderer requirements, and
 verification commands:
@@ -150,6 +151,7 @@ verification commands:
 - {doc}`Genesis <../2-user_guide/3-backends/5-genesis>`
 - {doc}`Newton <../2-user_guide/3-backends/7-newton>`
 - {doc}`Motrix <../2-user_guide/3-backends/2-motrix>`
+- {doc}`Drake <../2-user_guide/3-backends/6-drake>`
 
 ## Algorithm Extras
 

docs/sphinx/source/en/2-user_guide/3-backends/0-index.md

@@ -1,15 +1,15 @@
 # Simulation Backends
 
 The tensor-only Manager runtime currently exposes `mujoco`, `mjwarp`, `genesis`,
-`newton`, and `motrix`. User commands select one with `--sim`, which routes to
-the matching task owner YAML; do not switch a run by overriding
-`training.sim_backend` alone.
+`newton`, `motrix`, and the scoped Drake owner described below. User commands
+select one with `--sim`, which routes to the matching task owner YAML; do not
+switch a run by overriding `training.sim_backend` alone.
 
-The `drake`, `isaacgym`, `isaacsim`, and `superdex` adapters remain temporarily
-shelved by `unisim-core` during issue #1811. Their historical pages are retained
-for adapter context only and are not production support claims. The train/eval
-CLI rejects these names until new capability, parity, and support-matrix
-evidence is provided.
+The `isaacgym`, `isaacsim`, and `superdex` adapters remain temporarily shelved
+by `unisim-core` during issue #1811. Their historical pages are retained for
+adapter context only and are not production support claims. The train/eval CLI
+rejects these names until new capability, parity, and support-matrix evidence is
+provided.
 
 ## Runtime Prerequisites
 
@@ -24,6 +24,10 @@ evidence is provided.
 - Motrix requires the `motrix` extra and provides a CPU-authoritative packed
   HOST_BRIDGE; the current canonical support scope is SAC `g1_walk_flat` and
   FlashSAC `g1_motion_tracking`.
+- Drake requires the locally built DrakeUni batch extension. Its scoped PPO
+  `go2_joystick_flat` owner uses CPU physics and the packed HOST_BRIDGE
+  lifecycle; floating-root and reset-randomization events remain disabled until
+  the backend contract supports them.
 
 ## OS and GPU Support
 
@@ -34,6 +38,7 @@ evidence is provided.
 | Genesis | Linux x86_64 | Required: NVIDIA GPU and driver; only the `gs.gpu` channel is validated |
 | Newton | Linux | Required: NVIDIA CUDA; the selected-reset lane is device-resident |
 | Motrix | Linux / macOS / Windows | CPU-authoritative physics; Torch CUDA buffers are optional |
+| Drake | Linux x86_64 / Apple Silicon macOS | CPU-authoritative physics; Torch CUDA buffers are optional |
 
 Backend device requirements are independent of the learner device: MuJoCo can
 still train with its learner on CUDA, ROCm, MPS, or XPU. See the platform

docs/sphinx/source/en/2-user_guide/3-backends/6-drake.md

@@ -1,6 +1,8 @@
 # Drake Backend
 
-> **Drake is temporarily shelved.** This adapter is outside the tensor-only Manager runtime during #1811 and is not a production support claim. The train/eval CLI rejects it until new capability, parity, and support-matrix evidence is provided.
+> **Scoped runtime support.** Drake currently has one Manager-runtime owner:
+> PPO `go2_joystick_flat`. The page is not a broader production support claim.
+> Task support remains governed by the generated support matrix.
 
 
 Drake is an experimental CPU batch backend. UniLab still owns the task,
@@ -95,9 +97,10 @@ uv run train --algo ppo --task go2_joystick_flat --sim drake
 ```
 
 This uses the Drake owner configuration (`1024` environments, `151` iterations,
-and CPU training because Drake exposes float64 NumPy buffers). The Drake owner
-also uses the scene keyframe reset; floating-root randomization is not exposed
-by the current backend contract.
+and CPU training because Drake exposes a CPU-authoritative packed host bridge).
+The Drake owner uses the scene default/keyframe reset. Floating-root
+randomization and PD reset randomization are not exposed by the current backend
+contract.
 
 On Apple Silicon macOS, this command completed all 151 iterations locally
 (Drake 1.56.0, Python 3.13, 1024 environments) in about 254 seconds.

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

@@ -8,7 +8,7 @@ support beyond the evidence grade shown below.
 ## Backend Selection Rules
 
 - The tensor-only Manager runtime currently supports `mujoco`, `mjwarp`,
-  `genesis`, `newton`, and `motrix`.
+  `genesis`, `newton`, `motrix`, and the scoped Drake owner below.
 - The default backend is `mujoco`.
 - `--sim mjwarp` requires the `mjwarp` extra. Validated combinations are shown
   in the generated matrix below; all other entrypoints retain their matrix
@@ -19,12 +19,14 @@ support beyond the evidence grade shown below.
   are shown in the generated matrix.
 - `--sim motrix` requires the `motrix` extra; the CPU-authoritative packed
   HOST_BRIDGE profile is shown in the generated matrix.
+- `--sim drake` requires the local Drake extra and DrakeUni batch extension.
+  Its current scoped support claim is PPO `go2_joystick_flat` only.
 - `--algo`, `--task`, and `--sim` jointly select the owner YAML.
 - Do not treat `training.sim_backend` as a standalone backend switch.
-- `drake`, `isaacgym`, `isaacsim`, and `superdex` are temporarily outside this
-  runtime scope. Their adapters remain in UniSim, but they are not UniLab
-  production support claims; re-enabling requires capability, parity, and
-  support-matrix evidence (#1811).
+- `isaacgym`, `isaacsim`, and `superdex` are temporarily outside this runtime
+  scope. Their adapters remain in UniSim, but they are not UniLab production
+  support claims; re-enabling requires capability, parity, and support-matrix
+  evidence (#1811).
 
 ## Playback Differences
 
@@ -74,34 +76,35 @@ This table is derived from UniSim's SDK-free public static inventory. It describ
 |---|---|---|---|---|---|---|---|---|---|---|---|
 | `mujoco` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge only; no ROCm CUDA-only fallback | In-process; no external Python worker | unknown | unknown | Unsupported | Exact |
 | `motrix` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge only; no ROCm CUDA-only fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
+| `drake` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge only; no ROCm CUDA-only fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
 | `mjwarp` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
 | `newton` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
 | `genesis` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
 
 ### Entrypoint x Task Owner
 
-| Entrypoint | Task owner | MuJoCo | Motrix | mjwarp | Newton | Genesis |
-|------------|------------|---|---|---|---|---|
-| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | - | - | - |
-| PPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Tested | Registered | Configured |
-| PPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | - | - | - | - |
-| PPO (torch) | `x2_wall_flip_tracking` (X2 wall flip tracking) | Tested | - | - | - | - |
-| PPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | - | - | - |
-| PPO (torch) | `allegro_inhand_grasp` (allegro inhand grasp) | Tested | - | - | - | - |
-| PPO (torch) | `g1_box_tracking` (g1 box tracking) | Tested | - | - | - | - |
-| PPO (torch) | `stewart_balance` (stewart balance) | Tested | - | - | - | - |
-| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | - | - | - |
-| APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Registered | Registered | Registered |
-| APPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | - | - | - | - |
-| APPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | - | - | - |
-| 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 |
-| SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | - | - | - |
-| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | - | - | - |
-| 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 |
-| WarpSAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | Tested | Registered | Registered |
-| WarpSAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | Tested | Registered | Registered |
+| Entrypoint | Task owner | MuJoCo | Motrix | Drake | mjwarp | Newton | Genesis |
+|------------|------------|---|---|---|---|---|---|
+| PPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Configured | - | - | - |
+| PPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Tested | Registered | Configured |
+| PPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | - | - | - | - | - |
+| PPO (torch) | `x2_wall_flip_tracking` (X2 wall flip tracking) | Tested | - | - | - | - | - |
+| PPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | - | - | - | - |
+| PPO (torch) | `allegro_inhand_grasp` (allegro inhand grasp) | Tested | - | - | - | - | - |
+| PPO (torch) | `g1_box_tracking` (g1 box tracking) | Tested | - | - | - | - | - |
+| PPO (torch) | `stewart_balance` (stewart balance) | Tested | - | - | - | - | - |
+| APPO (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | Registered | - | - | - |
+| APPO (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Registered | Registered | Registered |
+| APPO (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | - | - | - | - | - |
+| APPO (torch) | `allegro_inhand` (Allegro in-hand) | Tested | - | - | - | - | - |
+| 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 |
+| SAC (torch) | `g1_wbt_obs` (g1 wbt obs) | Tested | - | - | - | - | - |
+| FlashSAC (torch) | `go2_joystick_flat` (Go2 joystick) | Tested | - | 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 |
+| WarpSAC (torch) | `g1_walk_flat` (G1 walk flat) | Tested | Registered | - | Tested | Registered | Registered |
+| WarpSAC (torch) | `g1_motion_tracking` (G1 motion tracking) | Tested | Registered | - | Tested | Registered | Registered |
 
 ### Source Index
 

docs/sphinx/source/zh_CN/1-getting_started/2-installation.md

@@ -108,8 +108,9 @@ Hugging Face endpoint 无法访问时，可设置 `HF_ENDPOINT=https://hf-mirror
 
 ## 后端 Extras
 
-当前 tensor-only Manager runtime 暴露五个生产后端：`mujoco`、`mjwarp`、
-`genesis`、`newton`、`motrix`。各自的仿真依赖均为可选。
+当前 tensor-only Manager runtime 暴露五个生产后端（`mujoco`、`mjwarp`、
+`genesis`、`newton`、`motrix`），另有一个 scoped Drake owner。各自的仿真依赖
+均为可选。
 
 ```bash
 # 一次安装全部 scoped 后端。
@@ -123,8 +124,9 @@ uv sync --extra mujoco --extra mjwarp --extra genesis --extra newton --extra mot
 | Genesis | `uv sync --extra genesis` | 已验证路径使用 Linux x86_64、NVIDIA GPU 及固定版本 torch/Genesis |
 | Newton | `uv sync --extra newton` | Linux CUDA，并钉定 newton / MuJoCo-Warp / Warp 版本 |
 | Motrix | `uv sync --extra motrix` | CPU-authoritative MotrixSim 与 packed Torch HOST_BRIDGE 传输 |
+| Drake | `make setup-drake` | 本地 Drake C++ 前缀与 DrakeUni native batch extension；当前 scoped 到 PPO `go2_joystick_flat` |
 
-`drake`、`isaacgym`、`isaacsim` 与 `superdex` 适配器在 issue #1811 期间由
+`isaacgym`、`isaacsim` 与 `superdex` 适配器在 issue #1811 期间由
 `unisim-core` 暂时搁置。它们的 extras 和历史后端页面不构成生产支持声明；
 在提供新的 capability、parity 与支持矩阵证据之前，UniLab train/eval CLI
 会直接拒绝这些后端。
@@ -136,6 +138,7 @@ runtime 变量、渲染器要求和验证命令见 scoped 后端页面：
 - {doc}`Genesis <../2-user_guide/3-backends/5-genesis>`
 - {doc}`Newton <../2-user_guide/3-backends/7-newton>`
 - {doc}`Motrix <../2-user_guide/3-backends/2-motrix>`
+- {doc}`Drake <../2-user_guide/3-backends/6-drake>`
 
 ## 算法 Extras
 

docs/sphinx/source/zh_CN/2-user_guide/3-backends/0-index.md

@@ -1,10 +1,11 @@
 # 仿真后端
 
 当前 tensor-only Manager runtime 暴露 `mujoco`、`mjwarp`、`genesis`、
-`newton` 与 `motrix`。用户命令通过 `--sim` 选择后端，并组合对应的 task owner
-YAML；不要单独覆盖 `training.sim_backend`。
+`newton`、`motrix`，以及下文描述的 scoped Drake owner。用户命令通过 `--sim`
+选择后端，并组合对应的 task owner YAML；不要单独覆盖
+`training.sim_backend`。
 
-`drake`、`isaacgym`、`isaacsim` 与 `superdex` 适配器在 issue #1811 期间由
+`isaacgym`、`isaacsim` 与 `superdex` 适配器在 issue #1811 期间由
 `unisim-core` 暂时搁置。历史页面仅用于 adapter 背景，不构成生产支持声明。
 在提供新的 capability、parity 与支持矩阵证据之前，train/eval CLI 会直接
 拒绝这些后端。
@@ -20,6 +21,10 @@ YAML；不要单独覆盖 `training.sim_backend`。
   为 SAC `g1_walk_flat` 和 FlashSAC `g1_motion_tracking`。
 - Motrix 需要 `motrix` extra，提供 CPU-authoritative packed HOST_BRIDGE；当前
   canonical 支持范围为 SAC `g1_walk_flat` 和 FlashSAC `g1_motion_tracking`。
+- Drake 需要本地编译的 DrakeUni batch extension。其 scoped PPO
+  `go2_joystick_flat` owner 使用 CPU physics 与 packed HOST_BRIDGE
+  lifecycle；在 backend contract 补齐前，浮动根重置与 reset randomization
+  事件保持禁用。
 
 ## OS 与 GPU 支持
 
@@ -30,6 +35,7 @@ YAML；不要单独覆盖 `training.sim_backend`。
 | Genesis | Linux x86_64 | 要求：NVIDIA GPU 与 driver；仅验证 `gs.gpu` channel |
 | Newton | Linux | 要求：NVIDIA CUDA；selected-reset 通道为 device-resident |
 | Motrix | Linux / macOS / Windows | CPU-authoritative physics；Torch CUDA buffer 可选 |
+| Drake | Linux x86_64 / Apple Silicon macOS | CPU-authoritative physics；Torch CUDA buffer 可选 |
 
 后端设备要求与 learner 设备独立：MuJoCo 仍可以让 learner 使用 CUDA、ROCm、MPS
 或 XPU。平台配置见 {doc}`../../1-getting_started/2-installation`。

docs/sphinx/source/zh_CN/2-user_guide/3-backends/6-drake.md

@@ -1,6 +1,8 @@
 # Drake 后端
 
-> **Drake 暂时搁置。** 该适配器在 #1811 期间不属于 tensor-only Manager runtime，也不构成生产支持声明。在提供新的 capability、parity 与支持矩阵证据之前，train/eval CLI 会直接拒绝该后端。
+> **Scoped runtime 支持。** Drake 当前只有一个 Manager-runtime owner：PPO
+> `go2_joystick_flat`。本页不构成更广泛的生产支持声明；任务支持范围以生成的
+> 支持矩阵为准。
 
 
 Drake 是实验性的 CPU 批量物理后端；task、reward、observation 和训练循环仍由
@@ -89,8 +91,9 @@ uv run train --algo ppo --task go2_joystick_flat --sim drake
 ```
 
 该命令使用 Drake owner 配置中的 1024 个环境、151 次迭代，并固定使用 CPU 训练
-（Drake 暴露的是 float64 NumPy buffer）。Drake owner 使用场景 keyframe 重置；当前
-backend contract 尚未提供浮动根状态随机化。只想验证安装时，可临时追加
+（Drake 暴露的是 CPU-authoritative packed host bridge）。Drake owner 使用场景默认
+/keyframe 重置；当前 backend contract 尚未提供浮动根状态与 PD reset
+randomization。只想验证安装时，可临时追加
 `algo.max_iterations=1 algo.num_envs=4 algo.num_steps_per_env=4 training.no_play=true
 env.drake_nthread=1`；这些不是生产参数。