TATP-233 · Oct 8, 2026 · Oct 8, 2026
@@ -75,16 +75,12 @@ src/unilab/assets/motions/g1/**/*.npz
 src/unilab/assets/motions/g1/**/*.csv
 src/unilab/assets/motions/g1/flip
 src/unilab/assets/motions/g1/flip_npz
-src/unilab/assets/motions/x2/*.npz
-src/unilab/assets/motions/x2/*.csv
 
 # Robot mesh assets (downloaded from HF at runtime)
 src/unilab/assets/bots/
-src/unilab/assets/robots/x2/meshes/*.STL
 src/unilab/assets/robots/g1/assets/
 src/unilab/assets/robots/g1/textures/
 src/unilab/assets/robots/go2/assets/
-src/unilab/assets/robots/allegro_hand/assets/
 src/unilab/assets/robots/sharpa_wave/meshes/
 
 # Grasp cache assets (downloaded from HF at runtime)

@@ -57,7 +57,7 @@ rewriting the task's environment lifecycle.
   becoming a CUDA-resident simulator. The learner can run on CUDA, ROCm, MPS,
   or XPU; the [support matrix](https://unilabsim.github.io/UniLab-doc/en/5-reference/5-support_matrix.html)
   records the evidence level of each backend/task combination.
-- **Accelerate replay-based off-policy training.** FastSAC/FlashSAC/WarpSAC lets
+- **Accelerate replay-based off-policy training.** FastSAC/FlashSAC lets
   simulation data collection overlap with learner updates. The paper reports
   3–10× end-to-end gains on representative configurations; see [Why UniLab](https://unilabsim.github.io/UniLab-doc/en/why_unilab.html)
   for scope and measurements.

@@ -51,7 +51,7 @@ learner，而无需重写任务的 environment 生命周期。
   不必先变成 CUDA-resident simulator，也可以向 accelerator learner 提供数据；learner
   可以运行在 CUDA、ROCm、MPS 或 XPU 上。每个 backend/task 组合的证据等级请查看
   [支持矩阵](https://unilabsim.github.io/UniLab-doc/zh_CN/5-reference/5-support_matrix.html)。
-- **加速 replay-based off-policy 训练。** FastSAC/FlashSAC/WarpSAC 让仿真数据采集与 learner
+- **加速 replay-based off-policy 训练。** FastSAC/FlashSAC 让仿真数据采集与 learner
   update 重叠。论文在代表性配置上报告了 3–10 倍端到端收益；测量范围和限制见
   [为什么选择 UniLab](https://unilabsim.github.io/UniLab-doc/zh_CN/why_unilab.html)。
 

@@ -15,7 +15,7 @@ orphan: true
 ## Context
 
 [Discussion #1541](https://github.com/Motphys/UniLab/discussions/1541)
-指出，legacy domain-randomization provider 与 SimToolReal 大量工具模型瓶颈有共同根因：
+指出，legacy domain-randomization provider 与大量工具模型瓶颈有共同根因：
 backend 缺少 per-env model identity / model-field indirection。因此 task 曾被迫在
 UniLab 侧编译多个 engine model，或在 reset 协议中扩展模型变更语义。
 
@@ -94,7 +94,7 @@ backend 名称、可选 package import、executor introspection 或异常降级
 
 - 让 task config 直接持有或返回 `MjSpec`。拒绝原因：task YAML 变成 MuJoCo-specific，
   UniSim 难以保持 MJWarp 兼容，且 public plan 不再 pickle-safe。
-- 在 UniLab 为每个 env 编译完整 model。拒绝原因：这正是 SimToolReal 内存和冷启动
+- 在 UniLab 为每个 env 编译完整 model。拒绝原因：这正是大量工具模型内存和冷启动
   瓶颈，并把 engine realization 上移到错误 owner。
 - 在 reset provider 中切换 model identity。拒绝原因：破坏派生常量、CUDA graph 和
   playback 的生命周期假设，也会延长 legacy DR 协议共存。

@@ -27,7 +27,7 @@ fail closed，也没有可审计的 runtime evidence。
 - 新增共享 off-policy owner 设置
   `training.cuda_process_sharing: null | mps`，默认 `null` 保持现状。
 - 初始有效范围限定为 Linux、NVIDIA CUDA、`training.sim_backend=mjwarp`、单主机、
-  `world_size=1` 的 SAC/FastSAC/FlashSAC/WarpSAC 共享路径。
+  `world_size=1` 的 SAC/FastSAC/FlashSAC 共享路径。
 - `mps` 是显式请求：所有前置条件在 env probe、learner construction 与 collector
   spawn 之前验证，失败时给出第一个未满足条件和启动既有 daemon 的命令，不 fallback。
 - rank-local 物理一致性按 GPU UUID 判断，而不是 CUDA ordinal。
@@ -69,7 +69,6 @@ fail closed，也没有可审计的 runtime evidence。
 - `src/unilab/scripts/train_offpolicy.py`
 - `src/unilab/conf/sac/config.yaml`
 - `src/unilab/conf/flashsac/config.yaml`
-- `src/unilab/conf/warpsac/config.yaml`
 - `tests/training/test_cuda_process_sharing.py`
 - `tests/algos/test_offpolicy_double_buffer_runner.py`
 - `docs/sphinx/source/en/2-user_guide/1-training/7-tensor_runtime_production.md`

@@ -3,8 +3,7 @@
 Concrete RL tasks split by family:
 
 - **locomotion** — Go2 and Unitree G1 reference owners
-- **manipulation** — Allegro in-hand cube and Stewart balance
-- **motion_tracking** — G1 and X2 whole-body motion tracking
+- **motion_tracking** — G1 whole-body motion tracking
 
 Every task is registered into the task `Registry` so it can be selected via
 `uv run train --algo <algo> --task <name> --sim <backend>`.
@@ -14,7 +13,6 @@ Unitree production variants are documented in `unitree_rl_unilab`.
 :maxdepth: 2
 
 locomotion
-manipulation
 motion_tracking
 ```
 

@@ -1,6 +1,6 @@
 # `unilab.tasks.motion_tracking`
 
-Whole-body motion-tracking tasks for G1 and X2 robots.
+Whole-body motion-tracking tasks for G1 robots.
 
 ```{eval-rst}
 .. autosummary::
@@ -10,5 +10,4 @@ Whole-body motion-tracking tasks for G1 and X2 robots.
 
    unilab.tasks.motion_tracking.common
    unilab.tasks.motion_tracking.g1
-   unilab.tasks.motion_tracking.x2
 ```
@@ -155,7 +155,6 @@ committed benchmark manifest or separate recommendation metadata.
 | --- | --- | --- |
 | Go2 joystick | PPO, FlashSAC | PPO has tested MuJoCo and Motrix rows. FlashSAC has MuJoCo owner YAMLs for `go2_joystick_flat`. |
 | G1 reference locomotion / tracking | PPO, APPO, SAC | PPO, APPO, and SAC include committed MuJoCo and Motrix owner YAMLs for retained G1 tasks. |
-| Allegro in-hand | PPO, APPO | PPO and APPO have committed MuJoCo and Motrix owner YAMLs for Allegro in-hand tasks. |
 
 Unitree production task families are maintained in `unitree_rl_unilab` and are
 not counted as core repository evidence.

@@ -42,9 +42,8 @@ Run a pre-trained policy before changing any task configuration:
 uv run demo dance
 ```
 
-Available demo names are `teaser`, `dance`, `wallflip`, `boxtracking`,
-and `inhandgrasp`. Use `uv run demo --help` for device and refresh
-options.
+Available demo names are `teaser` and `dance`. Use `uv run demo --help` for
+device and refresh options.
 
 ## Train A Task
 
@@ -75,14 +74,6 @@ community mirror before running training, eval, or demo commands:
 export HF_ENDPOINT=https://hf-mirror.com
 ```
 
-X2 robot meshes auto-download from Hugging Face (`unilabsim/unilab-robots`) into
-`src/unilab/assets/robots/x2/meshes/` on first run; no manual step is needed. To
-pre-fetch them:
-
-```bash
-uv run unilab-pull-assets --robot x2
-```
-
 On macOS, the CLI routes Motrix interactive playback through `mxpython` when
 needed. Use `--render-mode record` for headless video export or
 `--render-mode none` to skip rendering.

@@ -11,7 +11,6 @@ keeps the lower-level scripts available for debugging Hydra composition.
 | APPO | `uv run train --algo appo --task <task> --sim <backend>` | `src/unilab/scripts/train_appo.py` |
 | SAC | `uv run train --algo sac --task <task> --sim <backend>` | `src/unilab/scripts/train_sac.py` |
 | FlashSAC | `uv run train --algo flashsac --task <task> --sim <backend>` | `src/unilab/scripts/train_flashsac.py` |
-| WarpSAC | `uv run train --algo warpsac --task <task> --sim <backend>` | `src/unilab/scripts/train_warpsac.py` |
 
 Examples:
 
@@ -109,13 +108,10 @@ dependency; the `viser.*` config group (`viser.port`, `viser.max_envs`,
 
 ```bash
 uv run demo dance
-uv run demo wallflip
-uv run demo boxtracking
-uv run demo inhandgrasp
 uv run demo dance --refresh --device cpu
 ```
 
-Available demos: `teaser`, `dance`, `wallflip`, `boxtracking`, `inhandgrasp`.
+Available demos: `teaser`, `dance`.
 Each demo fetches a pre-trained checkpoint from the
 `unilabsim/unilab-checkpoints` Hugging Face dataset on first run and caches it
 under `src/unilab/assets/checkpoints/<demo>/model_0.pt`. Pass `--refresh` to

@@ -8,7 +8,7 @@ persistent backend, which keeps each unsmoothed iteration value.
 
 This page first covers the log directory shared by all algorithms and the
 Manager-Based reward metric contract, then documents the off-policy terminal used by
-SAC / FlashSAC / WarpSAC and APPO. Persisted backend tags follow the canonical schema
+SAC / FlashSAC and APPO. Persisted backend tags follow the canonical schema
 owned by `uni_rl.logging.metric_schema`: millisecond timing fields end in `_ms`,
 while `Perf/learning_time`, `Perf/collection_time`, and `Perf/iteration_time` are
 seconds. A few terminal fields (`Other`, phase percentages, collector cycle total,
@@ -31,7 +31,6 @@ stack overrides `training.log_root` or `training.log_dir`:
 | APPO | `logs/appo/<task>/` | `src/unilab/conf/appo/config.yaml` |
 | SAC | `logs/fast_sac/<task>/` | `src/unilab/conf/sac/config.yaml` |
 | FlashSAC | `logs/flash_sac/<task>/` | `src/unilab/conf/flashsac/config.yaml` |
-| WarpSAC | `logs/warp_sac/<task>/` | `src/unilab/conf/warpsac/config.yaml` |
 
 A run directory is named `YYYY-MM-DD_HH-MM-SS_<sim_backend>`, for example
 `2026-03-09_18-30-00_mujoco`. Common artifacts include `run_config.json`,

@@ -1,6 +1,6 @@
 # Tensor runtime
 
-SAC, FlashSAC, and WarpSAC resolve their spawn-facing tensor-runtime settings
+SAC and FlashSAC resolve their spawn-facing tensor-runtime settings
 before probing the environment or constructing a learner. The owner YAML values
 are collected into one bounded `TensorRuntimeSettings` object and passed to the
 off-policy runner. Invalid values, unsupported bounds, and impossible memory

@@ -10,7 +10,6 @@ lives, and which command shape selects it. For general flags, see
 | APPO | async on-policy | `src/unilab/scripts/train_appo.py` | `src/unilab/conf/appo/config.yaml` |
 | SAC | off-policy | `src/unilab/scripts/train_sac.py` | `src/unilab/conf/sac/config.yaml` |
 | FlashSAC | off-policy | `src/unilab/scripts/train_flashsac.py` | `src/unilab/conf/flashsac/config.yaml` |
-| WarpSAC | off-policy | `src/unilab/scripts/train_warpsac.py` | `src/unilab/conf/warpsac/config.yaml` |
 
 ```{toctree}
 :hidden:
@@ -19,5 +18,4 @@ lives, and which command shape selects it. For general flags, see
 2-appo
 3-sac
 4-flash_sac
-5-warpsac
 ```
@@ -26,7 +26,7 @@ provided.
   FlashSAC `g1_motion_tracking`.
 - SuperDex requires the `superdex` extra on CPython 3.12/3.13 Linux x86_64 and
   provides a CPU-authoritative packed HOST_BRIDGE; the current scope is the
-  configured Go2 and FR3 research owners.
+  configured Go2 research owner.
 - 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

@@ -1,29 +1,22 @@
 # SuperDex Backend
 
 > SuperDex participates in the tensor-only Manager runtime as a packed CPU
-> `HOST_BRIDGE` backend. Its Go2 and FR3 owners remain **Configured** research
+> `HOST_BRIDGE` backend. Its Go2 owner remains a **Configured** research
 > profiles, not full-training or cross-platform production claims.
 
 
 SuperDex is an optional CPU physics adapter owned by `unisim.backend.superdex`.
-The initial UniLab owner is the fixed-base `FR3JointTarget` task:
-`src/unilab/conf/ppo/task/fr3_joint_target/superdex.yaml`. It uses seven torque
-actions, 21 observation values, selected tensor joint-state resets, and
-`TorchEnv`. Its support level is **Configured**; bounded rollout or short
-training checks do not establish full-training performance or platform support.
-The implementation is tracked in [#1534](https://github.com/Motphys/UniLab/issues/1534)
-under [roadmap #1533](https://github.com/Motphys/UniLab/issues/1533).
+The checked-in owner is `go2_joystick_flat/superdex`. Its support level is
+**Configured**; bounded rollout or short training checks do not establish
+full-training performance or platform support.
 
 ## Installation
 
 SuperDex Physics/Robotics 1.0.0 is published as Python wheels and is an
 optional UniLab extra; no native source build is required. The wheels carry
 the native batch executor and support CPython 3.12/3.13 on Linux x86_64 only;
 on other platforms the extra is empty and the CLI reports a targeted runtime
-diagnostic. CPU physics does not require CUDA. The FR3 owner has no record
-(video) playback — the `.superdex_bot` asset carries no MJCF visual model — so
-play defaults to the native interactive viewer (`play_render_mode=interactive`
-with `play_env_num=1`); use `training.play_render_mode=none` for headless runs.
+diagnostic. CPU physics does not require CUDA.
 
 ```bash
 # Source checkout (default Python 3.13; wheels support CPython 3.12/3.13):
@@ -42,19 +35,7 @@ are a temporary unilabsim build (`superdex-physics-uni` /
 the `superdex-physics` / `superdex-robotics` wheels, UniSim switches the
 package names and UniLab needs no change.
 
-Native FR3 assets are hosted on Hugging Face
-([unilabsim/unilab-robots](https://huggingface.co/datasets/unilabsim/unilab-robots)),
-like the other robot mesh assets; the wheels do not carry robot binaries. The
-asset hub registers `bots/arms/fr3_v2/fr3_v2.superdex_bot` and downloads the
-snapshot into `src/unilab/assets/` on first use, checking its collision SDF,
-render files, `LICENSE` and `NOTICE` before constructing physics. To pre-fetch
-the assets (e.g. for CI or offline prep):
-
-```bash
-uv run unilab-pull-assets --robot fr3_v2
-```
-
-To audit a local `project_superdex` checkout instead, set
+To audit a local `project_superdex` checkout, set
 `SUPERDEX_ASSETS_PATH=/absolute/path/to/project_superdex/assets`, or set
 `env.superdex_assets_root=/absolute/path/to/project_superdex/assets` to
 override it for a specific owner invocation. An explicit root takes precedence
@@ -65,20 +46,6 @@ A local source build is only needed when changing the SuperDex engine itself:
 builds the native extensions and links the local UniSim/UniLab checkouts
 editable. Regular use does not need it.
 
-## Run the FR3 Task
-
-```bash
-uv run --no-sync train --algo ppo --task fr3_joint_target --sim superdex \
-  algo.max_iterations=2 algo.num_steps_per_env=16 \
-  algo.algorithm.num_learning_epochs=1
-```
-
-The target joint positions, rewards, reset ranges and action scales live in the
-task's `base.yaml`. The torque bounds `[20,20,20,20,5,5,5]` Nm are an explicit
-research profile, not rated hardware limits. `superdex_effort_limits` declares
-the same bounds at the native backend boundary. The SDK remains single-threaded;
-the native scene executor below owns all supported CPU parallelism.
-
 ## Run the Go2 Task
 
 The `go2_joystick_flat/superdex` owner trains and evaluates the Go2 quadruped
@@ -213,17 +180,6 @@ through the production playback session and executes 64 SuperDex control steps
 without a renderer. It checks finite values and interface compatibility; a
 two-iteration checkpoint is not expected to walk reliably.
 
-```bash
-uv run --no-sync pytest tests/assets/test_superdex_assets.py \
-  tests/envs/test_fr3_superdex.py tests/test_cli_runtime_requirements.py -q
-```
-
-The optional native tests require the SDK and the FR3 assets (downloaded from
-Hugging Face on demand, or provided through `SUPERDEX_ASSETS_PATH`); they cover
-finite rollout data, selected reset isolation, immediate observation refresh
-and a spawned `EnvFactory`. Missing runtime/assets produce an explicit skip;
-such a run is not native validation. The base asset/config tests need no native
-asset checkout.
 
 Engine conversion and physics live in UniSim; asset registration, Hydra and task
 terms remain in UniLab. See

@@ -15,13 +15,7 @@ Go2 and G1 core reference owners; Unitree production variants are downstream.
 :::{grid-item-card} Motion tracking
 :link: 2-motion_tracking
 :link-type: doc
-G1/X2 motion-tracking reference profiles.
-:::
-
-:::{grid-item-card} Manipulation
-:link: 3-manipulation
-:link-type: doc
-Allegro in-hand rotation and grasp generation.
+G1 motion-tracking reference profiles.
 :::
 
 ::::
@@ -31,5 +25,4 @@ Allegro in-hand rotation and grasp generation.
 
 1-locomotion
 2-motion_tracking
-3-manipulation
 ```