Feat/a2arm pos force

.github/CODEOWNERS

@@ -7,8 +7,8 @@
 /scripts/train_*.py @TATP-233 @caozx1110
 
 # Task and environment ownership
-/src/unilab/envs/motion_tracking/ @caozx1110
-/src/unilab/envs/manipulation/ @Mingrui-Yu
+/src/unilab/tasks/motion_tracking/ @caozx1110
+/src/unilab/tasks/manipulation/ @Mingrui-Yu
 /scripts/motion/ @caozx1110
 
 # Project process and docs

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -37,7 +37,7 @@ body:
       label: Reproduction
       description: Exact command, config, or sequence that reproduces the issue.
       placeholder: |
-        uv run scripts/train_offpolicy.py algo=sac task=sac/g1_walk_flat/mujoco ...
+        uv run scripts/train_sac.py task=g1_walk_flat/mujoco ...
     validations:
       required: true
 

.gitignore

@@ -58,7 +58,7 @@ run_summary.json
 src/unilab/assets/checkpoints/
 
 scripts/benchmark/outputs/
-src/unilab/algos/torch/rsl_rl
+src/unilab/algos/rsl_rl
 
 third-party
 temp/
@@ -81,6 +81,12 @@ src/unilab/assets/motions/x2/*.csv
 
 # Robot mesh assets (downloaded from HF at runtime)
 src/unilab/assets/robots/x2/meshes/*.STL
+src/unilab/assets/robots/a2arm/meshes/*
+!src/unilab/assets/robots/a2arm/meshes/.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
 
 # Grasp cache assets (downloaded from HF at runtime)
 src/unilab/assets/caches/*.npy

AGENTS.md

@@ -39,27 +39,27 @@ UniLab 是一个 **高性能、模块化、contract 驱动** 的 RL infrastructu
 
 ## Sim2Sim 跨后端配置契约
 
-`src/unilab/training/sim2sim.py` 按 dotted path 维护三类字段：
+`src/unilab/utils/sim2sim.py` 按 dotted path 维护三类字段：
 
 - **DENYLIST**（差异即 `CrossBackendIncompatibleError`）：`algo.obs_groups`、`env.control_config.action_scale`、`algo.policy.actor_hidden_dims` / `critic_hidden_dims`、`algo.empirical_normalization` / `algo.obs_normalization`、`env.sampling_mode`。`env.*` 子集对**任一方向**的不对称出现也 fail-closed；`algo` 专属字段目标缺省时按设计跳过（跨算法合法）。
 - **WARNING_LIST**：`reward.*`、`env.control_config.simulate_action_latency`、`env.ctrl_dt`。
 - **ALLOWLIST**（自由覆盖）：`training.sim_backend`、`env.scene`、`training.play_steps`、`env.domain_rand`、`env.noise_config`、`env.commands.vel_limit`。
 
-训练时 `ExperimentTracker.start()` 把上述字段写入 `run_config.json` 的 `contract_snapshot`（不改 checkpoint 格式，旧 run 无 snapshot 时 fallback + warning）；五个 play 入口在建 env 前调用 `resolve_sim2sim_config` 校验，并用 `policy_load_dim_guard` 包裹 checkpoint 加载以把维度不匹配的隐晦报错重抛为显式诊断。设 `training.sim2sim_strict=false` 可把 DENYLIST 差异降级为 warning（默认 `true`）。DENYLIST 字段在每个后端 owner 配置中显式声明并保持跨后端一致（范例：`conf/ppo/task/g1_walk_flat/{mujoco,motrix}.yaml`）；跨后端契约审计见 `scripts/audit_sim2sim_contracts.py`。
+训练时 `ExperimentTracker.start()` 把上述字段写入 `run_config.json` 的 `contract_snapshot`（不改 checkpoint 格式，旧 run 无 snapshot 时 fallback + warning）；五个 play 入口在建 env 前调用 `resolve_sim2sim_config` 校验，并用 `policy_load_dim_guard` 包裹 checkpoint 加载以把维度不匹配的隐晦报错重抛为显式诊断。设 `training.sim2sim_strict=false` 可把 DENYLIST 差异降级为 warning（默认 `true`）。DENYLIST 字段在共享 base owner 与后端 owner 配置中显式声明并保持跨后端一致（范例：`conf/ppo/task/g1_walk_flat/{base,mujoco,motrix}.yaml`）；跨后端契约审计见 `scripts/audit_sim2sim_contracts.py`。
 
 ## Pointers
 
 - PPO: `scripts/train_rsl_rl.py`
 - APPO: `scripts/train_appo.py`
-- SAC / TD3: `scripts/train_offpolicy.py`
+- SAC / TD3 / FlashSAC: `scripts/train_sac.py` / `scripts/train_td3.py` / `scripts/train_flashsac.py`
 - env contract: `src/unilab/base/np_env.py`
 - backend contract: `src/unilab/base/backend/base.py`
 - training run helpers: `src/unilab/training/run.py`
 - visualization helpers: `src/unilab/visualization/`
 - shared numeric helpers: `src/unilab/utils/rotation.py`, `src/unilab/utils/geometry.py`
 - config schema: `src/unilab/structured_configs.py`
 - async runner: `src/unilab/ipc/async_runner.py`
-- sim2sim 跨后端契约: `src/unilab/training/sim2sim.py`
+- sim2sim 跨后端契约: `src/unilab/utils/sim2sim.py`
 
 ## GitHub CLI (gh) 速查
 

Makefile

@@ -88,7 +88,7 @@ clean:
 	find . -type d -name ".ruff_cache" -exec rm -rf {} +
 	find . -type d -name "htmlcov" -exec rm -rf {} +
 	find . -type f -name ".coverage" -delete
-	rm -f train_appo.log train_offpolicy.log train_rsl_rl.log MUJOCO_LOG.TXT
+	rm -f train_appo.log train_sac.log train_flashsac.log train_rsl_rl.log MUJOCO_LOG.TXT
 	find src/unilab/assets/.cache -type f ! -name '.gitkeep' -delete 2>/dev/null || true
 	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

conf/appo/task/allegro_inhand/base.yaml

@@ -0,0 +1,138 @@
+# @package _global_
+# Canonical Allegro rotation Manager-Based task declaration. Backend leaves own
+# only backend identity and algorithm/runtime tuning.
+env:
+  scene:
+    model_file: src/unilab/assets/robots/allegro_hand/scene.xml
+    default_keyframe_name: home
+    entities:
+      robot:
+        root_body_name: ball
+        joint_names:
+          - ffj0
+          - ffj1
+          - ffj2
+          - ffj3
+          - mfj0
+          - mfj1
+          - mfj2
+          - mfj3
+          - rfj0
+          - rfj1
+          - rfj2
+          - rfj3
+          - thj0
+          - thj1
+          - thj2
+          - thj3
+        body_names: [ball, ff_tip, mf_tip, rf_tip, th_tip]
+        actuator_names:
+          - ffa0
+          - ffa1
+          - ffa2
+          - ffa3
+          - mfa0
+          - mfa1
+          - mfa2
+          - mfa3
+          - rfa0
+          - rfa1
+          - rfa2
+          - rfa3
+          - tha0
+          - tha1
+          - tha2
+          - tha3
+  sim_dt: 0.005
+  ctrl_dt: 0.05
+  max_episode_seconds: 20.0
+  observations:
+    policy:
+      history_length: 3
+      flatten_history_dim: true
+      terms:
+        rotation:
+          func: unilab.tasks.manipulation.allegro_inhand.manager_terms.AllegroRotationObservation
+          params:
+            entity_name: robot
+            action_name: hand
+            joint_noise: 0.02
+            torque_estimate_kp: 1.0
+            torque_estimate_kd: 0.1
+  actions:
+    hand:
+      _target_: unilab.tasks.manipulation.allegro_inhand.manager_terms.AllegroIncrementalPositionActionCfg
+      entity_name: robot
+      actuator_names: [".*"]
+      action_scale: 0.041666666666666664
+      raw_action_clip: [-1.0, 1.0]
+  events:
+    reset_scene_to_default:
+      func: unilab.envs.mdp.reset_scene_to_default
+      mode: reset
+    reset_hand_ball:
+      func: unilab.tasks.manipulation.allegro_inhand.manager_terms.AllegroHandBallReset
+      mode: reset
+      params:
+        entity_name: robot
+        # null explicitly selects the model home pose. A configured path is
+        # fail-closed when missing or malformed.
+        grasp_cache_path: null
+        joint_noise: 0.0
+        ball_velocity_noise: 0.0
+        ball_z_offset: 0.0
+    pd_gains:
+      func: unilab.envs.mdp.pd_gains
+      mode: reset
+      params:
+        kp_range: [1.0, 1.0]
+        kd_range: [0.1, 0.1]
+        operation: abs
+  terminations:
+    dropped:
+      func: unilab.tasks.manipulation.allegro_inhand.manager_terms.AllegroDropTermination
+      params:
+        observation_group: policy
+        observation_term: rotation
+        minimum_ball_height: 0.125
+    time_out:
+      func: unilab.envs.mdp.time_out
+      time_out: true
+  scale_rewards_by_dt: true
+  policy_observation_group: policy
+  critic_observation_group: null
+
+reward:
+  rotate:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.AllegroRotateReward
+    weight: 1.25
+    params:
+      state_term_name: dropped
+      rotation_axis: [0.0, 0.0, 1.0]
+      clip_min: -0.5
+      clip_max: 0.5
+  obj_linvel:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.object_linear_velocity_l1
+    weight: -0.3
+    params:
+      state_term_name: dropped
+  pose_diff:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.hand_pose_deviation_l2
+    weight: -0.3
+    params:
+      state_term_name: dropped
+  torque:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.estimated_torque_l2
+    weight: -0.1
+    params:
+      state_term_name: dropped
+  work:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.estimated_work_l2
+    weight: -2.0
+    params:
+      state_term_name: dropped
+  drop:
+    func: unilab.tasks.manipulation.allegro_inhand.manager_terms.dropped
+    weight: 0.0
+    params:
+      state_term_name: dropped

conf/appo/task/allegro_inhand/drake.yaml

@@ -0,0 +1,50 @@
+# @package _global_
+defaults:
+  - /task/allegro_inhand/base
+  - _self_
+
+training:
+  task_name: AllegroInhandRotation
+  sim_backend: drake
+  play_steps: 200
+  render_spacing: 0.5
+  cam_distance: 1.5
+  cam_lookat: [0.75, 0.75, 0]
+  cam_elevation: -20.0
+  replay_queue_size: 4
+algo:
+  num_envs: 1024
+  steps_per_env: 8
+  max_iterations: 3000
+  save_interval: 500
+  algorithm:
+    value_loss_coef: 4.0
+    entropy_coef: 0.01
+    learning_rate: 0.001
+    desired_kl: 0.025
+    adaptive_kl_factor: 2.0
+    adaptive_lr_factor: 1.5
+    num_learning_epochs: 5
+    num_mini_batches: 4
+    clip_param: 0.2
+    gamma: 0.99
+    lam: 0.95
+    max_grad_norm: 1.0
+    use_clipped_value_loss: true
+    schedule: adaptive
+  actor:
+    hidden_dims: [512, 256, 128]
+    activation: elu
+    obs_normalization: true
+    distribution_cfg:
+      class_name: rsl_rl.modules.distribution.GaussianDistribution
+      init_std: 1.0
+      std_type: scalar
+  critic:
+    hidden_dims: [512, 256, 128]
+    activation: elu
+    obs_normalization: true
+env:
+  drake_backend_mode: batch
+  events:
+    pd_gains: null

conf/appo/task/allegro_inhand/motrix.yaml

@@ -1,4 +1,8 @@
 # @package _global_
+defaults:
+  - /task/allegro_inhand/base
+  - _self_
+
 training:
   task_name: AllegroInhandRotation
   sim_backend: motrix
@@ -7,6 +11,9 @@ training:
   cam_distance: 1.5
   cam_lookat: [0.75, 0.75, 0]
   cam_elevation: -20.0
+env:
+  events:
+    pd_gains: null
 algo:
   num_envs: 16384
   steps_per_env: 8
@@ -39,27 +46,3 @@ algo:
     hidden_dims: [512, 256, 128]
     activation: elu
     obs_normalization: true
-reward:
-  scales:
-    rotate: 1.25
-    obj_linvel: -0.3
-    pose_diff: -0.3
-    torque: -0.1
-    work: -2.0
-    drop: 0.0
-  angvel_clip_min: -0.5
-  angvel_clip_max: 0.5
-  reset_z_threshold: 0.125
-env:
-  gen_grasp: false
-  max_episode_seconds: 20.0
-  grasp_cache_path: caches/allegro_grasp_50k.npy
-  # Keep only grasp/pose reset variation. All online DR terms stay disabled.
-  domain_rand:
-    randomize_base_mass: false
-    random_com: false
-    randomize_gravity: false
-    push_robots: false
-    joint_noise: 0.0
-    ball_vel_noise: 0.0
-    ball_z_offset: 0.0

conf/appo/task/allegro_inhand/mujoco.yaml

@@ -1,4 +1,8 @@
 # @package _global_
+defaults:
+  - /task/allegro_inhand/base
+  - _self_
+
 training:
   task_name: AllegroInhandRotation
   sim_backend: mujoco
@@ -41,27 +45,3 @@ algo:
     hidden_dims: [512, 256, 128]
     activation: elu
     obs_normalization: true
-reward:
-  scales:
-    rotate: 1.25
-    obj_linvel: -0.3
-    pose_diff: -0.3
-    torque: -0.1
-    work: -2.0
-    drop: 0.0
-  angvel_clip_min: -0.5
-  angvel_clip_max: 0.5
-  reset_z_threshold: 0.125
-env:
-  gen_grasp: false
-  max_episode_seconds: 20.0
-  grasp_cache_path: caches/allegro_grasp_50k.npy
-  # Keep only grasp/pose reset variation. All online DR terms stay disabled.
-  domain_rand:
-    randomize_base_mass: false
-    random_com: false
-    randomize_gravity: false
-    push_robots: false
-    joint_noise: 0.0
-    ball_vel_noise: 0.0
-    ball_z_offset: 0.0