feat(task): add RLinf Franka VLA pick-place deployment

.agents/skills/review-pr/SKILL.md

@@ -74,6 +74,56 @@ impact are established.
   file or subsystem coverage ledger and review dependency foundations before
   their consumers.
 
+## Explicitly requested PR publication
+
+Keep review read-only unless the user explicitly asks to publish the findings
+to a remote PR. A request to review, audit, or approve a change alone does not
+authorize remote comments or a review event.
+
+When publication is explicitly requested:
+
+1. Verify the repository, PR number, base branch, head commit, and changed
+   files before posting. If the target is not unambiguous, ask for the PR
+   identifier rather than guessing.
+2. Write every comment and review summary posted to the remote PR in English.
+   The conversational review may use the user's language, but translate the
+   finding faithfully before publication. Preserve code identifiers, paths,
+   and short log excerpts when they are needed as evidence.
+3. Publish each actionable finding as an inline comment on the smallest
+   relevant changed line when the hosting integration supports it. Put
+   findings that cannot be anchored to a changed line in a top-level review
+   summary. Preserve the finding priority and evidence in the posted text.
+4. Use a normal comment review by default. Submit `request changes` or
+   `approve` only when the user explicitly requests that exact review event;
+   never infer approval authority from a review request.
+5. Use an available GitHub connector or authenticated `gh` workflow. Do not
+   claim that comments were posted if the integration is unavailable or a
+   request fails. If publication is unavailable, return the complete,
+   copy-ready Markdown review instead.
+6. Report the published comment or review URLs, any findings that were not
+   posted, and any partial failures. Avoid duplicate comments when the same
+   finding has already been posted for the same head commit.
+
+Do not publish speculative open questions or residual risks as defects. They
+may be included in the top-level summary only when clearly labeled as such.
+
+### Trigger phrases
+
+The skill may be selected for review requests containing terms such as
+`review`, `audit`, `inspect`, `assess`, or `approve`. These terms authorize
+analysis only; they do not authorize remote PR changes.
+
+Publication requires an explicit remote-posting request, for example:
+
+- “publish/post/submit the review findings to the PR”
+- “add inline comments to the PR”
+- “leave the review comments on GitHub”
+- “把评审意见提交到 PR” or “在 PR 上添加评审评论”
+
+Phrases such as “review this PR”, “approve this PR”, or “给我评审建议” do not
+by themselves trigger publication. In particular, “approve this PR” selects
+the review task but does not authorize an `approve` review event.
+
 ## 1. Resolve the review target
 
 Read the applicable `AGENTS.md` instructions first. Determine the exact delta

docs/source/api_reference/public_api.rst

@@ -341,6 +341,7 @@ embodichain.lab.gym.envs.managers.actions
    QposTerm
    QposDenormalizedTerm
    QposNormalizedTerm
+   EefPoseGripperTerm
    EefPoseTerm
    QvelTerm
    QfTerm
@@ -2190,6 +2191,15 @@ embodichain_tasks.configs
 
    get_config_path
 
+embodichain_tasks.manipulation.repeated_pick_place
+------------------------------------------------
+
+.. currentmodule:: embodichain_tasks.manipulation.repeated_pick_place
+
+.. autosummary::
+
+   RepeatedPickPlaceRlinfEnv
+
 embodichain_tasks.manipulation.push_cube
 ----------------------------------------
 

embodichain/lab/gym/envs/managers/actions.py

@@ -42,6 +42,7 @@
 
 from embodichain.lab.sim.types import EnvAction
 from embodichain.utils.math import matrix_from_euler, matrix_from_quat
+
 from .action_manager import ActionTerm
 from .cfg import ActionTermCfg
 
@@ -54,12 +55,13 @@
 __all__ = [
     "DefaultJointPositionTerm",
     "DeltaQposTerm",
-    "QposTerm",
+    "EefPoseGripperTerm",
+    "EefPoseTerm",
+    "QfTerm",
     "QposDenormalizedTerm",
     "QposNormalizedTerm",
-    "EefPoseTerm",
+    "QposTerm",
     "QvelTerm",
-    "QfTerm",
 ]
 
 
@@ -259,6 +261,7 @@ def __init__(self, cfg: ActionTermCfg, env: EmbodiedEnv):
         super().__init__(cfg, env)
         self._scale = cfg.params.get("scale", 1.0)
         self._pose_dim = cfg.params.get("pose_dim", 7)  # 6 for euler, 7 for quat
+        self._part_name = cfg.params.get("part_name")
 
     @property
     def input_key(self) -> str:
@@ -286,21 +289,68 @@ def process_action(self, action: torch.Tensor) -> EnvAction:
                 f"EEF pose action must be 6D or 7D, got {scaled.shape[-1]}D"
             )
         # Batch IK: robot.compute_ik supports (num_envs, 4, 4) pose and (num_envs, dof) seed
+        joint_seed = current_qpos
+        part_joint_ids = None
+        if self._part_name is not None:
+            part_joint_ids = self._env.robot.get_joint_ids(
+                self._part_name, remove_mimic=True
+            )
+            joint_seed = current_qpos[:, part_joint_ids]
         ret, qpos_ik = self._env.robot.compute_ik(
             pose=target_pose,
-            joint_seed=current_qpos,
-        )
-        # Fallback to current_qpos where IK failed
-        result_qpos = torch.where(
-            ret.unsqueeze(-1).expand_as(qpos_ik), qpos_ik, current_qpos
+            joint_seed=joint_seed,
+            name=self._part_name,
         )
+        if part_joint_ids is None or qpos_ik.shape[-1] == current_qpos.shape[-1]:
+            result_qpos = torch.where(
+                ret.unsqueeze(-1).expand_as(qpos_ik), qpos_ik, current_qpos
+            )
+        else:
+            result_qpos = current_qpos.clone()
+            selected_qpos = torch.where(
+                ret.unsqueeze(-1).expand_as(qpos_ik),
+                qpos_ik,
+                current_qpos[:, part_joint_ids],
+            )
+            result_qpos[:, part_joint_ids] = selected_qpos
         return TensorDict(
             {"qpos": result_qpos, "ik_success": ret},
             batch_size=[batch_size],
             device=self.device,
         )
 
 
+class EefPoseGripperTerm(EefPoseTerm):
+    """Convert a 7D Cartesian action into arm IK targets and gripper joints.
+
+    The action is ``[x, y, z, roll, pitch, yaw, gripper]``. The first six
+    values use the same absolute-pose convention as :class:`EefPoseTerm`; the
+    final value is normalized from ``[-1, 1]`` to the active hand joint limits.
+    """
+
+    @property
+    def action_dim(self) -> int:
+        return 7
+
+    def process_action(self, action: torch.Tensor) -> EnvAction:
+        if action.shape[-1] != 7:
+            raise ValueError(
+                f"EefPoseGripperTerm expects 7D actions, got {action.shape[-1]}D."
+            )
+
+        arm_action = action[..., :6]
+        result = super().process_action(arm_action)
+        hand_joint_ids = self._env.robot.get_joint_ids("hand", remove_mimic=True)
+        if not hand_joint_ids:
+            raise ValueError("EefPoseGripperTerm requires a robot hand control part.")
+
+        limits = self._env.robot.body_data.qpos_limits[0, hand_joint_ids]
+        gripper = action[..., 6:7].clamp(-1.0, 1.0)
+        hand_qpos = limits[:, 0] + (gripper + 1.0) * 0.5 * (limits[:, 1] - limits[:, 0])
+        result["qpos"][:, hand_joint_ids] = hand_qpos.expand(-1, len(hand_joint_ids))
+        return result
+
+
 class QvelTerm(ActionTerm):
     """Joint velocity action: scale * action -> qvel.
 

embodichain_tasks/configs/components/embodiments/franka_panda_vla.yaml

@@ -0,0 +1,45 @@
+embodiment_id: franka_panda_vla
+
+simulation:
+  class_type: FrankaPanda
+  robot_type: panda
+  uid: FrankaPanda
+  solver_cfg:
+    arm:
+      class_type: PytorchSolver
+      end_link_name: fr3_hand_tcp
+      root_link_name: base
+      tcp:
+        - [1.0, 0.0, 0.0, 0.0]
+        - [0.0, 1.0, 0.0, 0.0]
+        - [0.0, 0.0, 1.0, 0.0]
+        - [0.0, 0.0, 0.0, 1.0]
+      num_samples: 30
+  link_attrs:
+    newton_gripper_contacts:
+      link_names_expr: ["fr3_leftfinger|fr3_rightfinger"]
+      attrs:
+        collision_props:
+          backend: newton
+          condim: 4
+        material_props:
+          backend: newton
+          ke: 40000.0
+          kd: 400.0
+          torsional_friction: 0.1
+          rolling_friction: 0.01
+
+sensor:
+  - sensor_type: Camera
+    uid: cam_high
+    width: 640
+    height: 480
+    intrinsics: [488.1665, 488.1665, 320.0, 240.0]
+    extrinsics:
+      eye: [-1.2, -1.2, 1.25]
+      target: [-0.1, 0.15, 0.20]
+      up: [0.0, 0.0, 1.0]
+
+skill_profile:
+  contract_id: single_arm_parallel_gripper
+  profile_id: franka_panda

embodichain_tasks/configs/tasks/manipulation/repeated_pick_place/env.rlinf.yaml

@@ -0,0 +1,54 @@
+environment_id: repeated_pick_place_rlinf
+physics: default
+max_episodes: 1
+max_episode_steps: 240
+num_envs: 1
+arena_space: 2.5
+
+simulation:
+  light:
+    direct:
+      - uid: main_light
+        light_type: sun
+        color: [0.6, 0.6, 0.6]
+        intensity: 5.0
+        direction: [0.0, 0.0, -1.0]
+  background:
+    - uid: target
+      shape:
+        shape_type: Cube
+        size: [0.11, 0.11, 0.01]
+      body_type: kinematic
+      init_pos: [-0.40, 0.48, 0.005]
+      attrs:
+        material_props: {static_friction: 0.0, dynamic_friction: 0.0}
+  rigid_object:
+    - uid: cube
+      shape:
+        shape_type: Cube
+        size: [0.05, 0.05, 0.05]
+      body_type: dynamic
+      init_pos: [-0.42, -0.08, 0.025]
+      attrs:
+        mass_props: {mass: 0.05}
+        rigid_props: {linear_damping: 0.2, angular_damping: 0.2}
+        material_props: {dynamic_friction: 0.97, static_friction: 0.99}
+
+env:
+  sim_steps_per_control: 4
+  observations:
+    eef_pose:
+      func: get_robot_eef_pose
+      mode: add
+      name: robot/eef_pose
+      params: {part_name: arm, position_only: false}
+  actions:
+    eef_pose_gripper:
+      func: EefPoseGripperTerm
+      params: {scale: 1.0, pose_dim: 6, part_name: arm}
+  rewards:
+    success_bonus:
+      func: success_reward
+      mode: add
+      weight: 10.0
+      params: {}

embodichain_tasks/configs/tasks/manipulation/repeated_pick_place/task.franka.rlinf.yaml

@@ -0,0 +1,7 @@
+id: RepeatedPickPlaceRlinf-Franka-v1
+
+environment:
+  component: env.rlinf.yaml
+
+embodiment:
+  component: ../../../components/embodiments/franka_panda_vla.yaml

embodichain_tasks/configs/tasks/manipulation/repeated_pick_place/task.franka.rlinf_expert.yaml

@@ -0,0 +1,14 @@
+id: TaskProgramRepeatedPickPlaceRlinfExpert-Franka-v1
+
+environment:
+  component: env.yaml
+
+task_program:
+  program: task_program/program.yaml
+  integration: task_program/integration.yaml
+  execution_policy: ../../../components/execution_policies/trajectory_open_loop.yaml
+
+embodiment:
+  component: ../../../components/embodiments/franka_panda_vla.yaml
+  overrides:
+    init_rot: [0.0, 0.0, 180.0]