embodichain.lab.task_program.integrations.simulation.workspace

embodichain.lab.task_program.integrations.simulation.workspace#

This compatibility adapter delegates workspace sampling and manipulation-pose checks to embodichain.lab.sim.scene.workspace, preserving the Task Program ValidationResult return type. Independent scene consumers use the canonical scene module and its SceneCheckReport. Both require one environment row and an explicit control part.

`RobotSceneWorkspace`	Preserve the Task Program import and its ValidationResult contract.
`SceneWorkspaceCandidate`	An object proposal derived from one valid cached joint configuration.

Object and manipulation frames#

Let T_object_tcp map TCP coordinates into the object frame. The object_grasp_pose parameter uses this convention. Sampling derives:

T_arena_object = T_arena_tcp @ inverse(T_object_tcp)
T_arena_tcp = T_arena_object @ T_object_tcp

All resulting poses use the local Z-up arena, excluding replication offsets. Robot.sample_reachable_pose already uses FK with the current base; callers must not apply that base transform or a Scene Engine coordinate conversion again.

Sampling and pose checks#

sample_object_poses uses a device-local seeded generator, preserves the workspace validity mask, and returns only valid cache entries. Optional bounds apply to object origins after the grasp transform is inverted. Exhausting the budget returns fewer proposals. Cache/configuration errors propagate rather than being hidden as ordinary rejections.

Each SceneWorkspaceCandidate retains its immutable pose proposal, joint seed, cache index, and optional score. Scores are available to the caller; this method does not rank or certify the proposals. Sampling does not establish support, stable placement, or collision-free motion.

check_object_pose composes the actual object and local grasp transforms, then calls the selected control part’s pose IK. A successful solver flag also requires finite joints within the selected part’s current limits and FK matching the requested position and orientation. This includes position-only solvers. position_tolerance_m defaults to 1e-3 and orientation_tolerance_rad to 1e-2; both must be finite and positive. Measured residuals are included in passing and FK-rejected checks. It also works for poses absent from the sparse cache. The returned mandatory workspace.pose_ik check applies to this manipulation target only. Changing an object’s height, orientation, or settled position requires another check, and grasp/pre-grasp/lift/place/retreat targets require their own checks. Motion between those poses and measured task success remain separate validation stages.

Producing candidates for the shared source#

The grasp transform is supplied from the application’s object affordance:

import torch

from embodichain.lab.sim.scene.expansion import (
    FixedSceneCandidateSource,
    SceneExpansionCfg,
)
from embodichain.lab.task_program.integrations.simulation.workspace import (
    RobotSceneWorkspace,
)

def propose_cube_layouts(
    workspace: RobotSceneWorkspace,
    object_grasp_pose: torch.Tensor,
) -> FixedSceneCandidateSource:
    candidates = workspace.sample_object_poses(
        "cube", object_grasp_pose, num_samples=3, seed=7, max_attempts=64
    )
    return FixedSceneCandidateSource(
        SceneExpansionCfg(seed=7, max_candidates=3, movable_entity_ids=("cube",)),
        parent_scene_id="pick-place-scene:revision-1",
        candidates=((candidate.pose_change,) for candidate in candidates),
        metadata=(("source", "robot-workspace"),),
    )

Construct the adapter as RobotSceneWorkspace(env.robot, control_part="arm") using the environment’s actual configured control-part name. After physical preparation, use the settled object’s pose in check_object_pose and combine that result with task-specific initial-state checks. Evaluate each proposal through embodichain.lab.task_program.integrations.scene_expansion and the concrete scene host; never replace measured acceptance with workspace success.

Public interfaces#

class embodichain.lab.task_program.integrations.simulation.workspace.RobotSceneWorkspace[source]#

Preserve the Task Program import and its ValidationResult contract.

New scene consumers should import the reusable service from embodichain.lab.sim.scene.workspace.

Methods:

check_object_pose(pose_change, ...[, joint_seed])

Convert scene checks to the existing execution validation contract.

check_object_pose(pose_change, object_grasp_pose, *, joint_seed=None)[source]#

Convert scene checks to the existing execution validation contract.

Parameters:

pose_change (ScenePoseChange) – Object pose in the local arena frame.
object_grasp_pose (Tensor | Sequence[Sequence[float]]) – TCP-to-object transform.
joint_seed (Sequence[float] | Tensor | None) – Optional selected-control-part joint seed.

Return type:

ValidationResult

Returns:

Required Task Program validation evidence.

class embodichain.lab.task_program.integrations.simulation.workspace.SceneWorkspaceCandidate[source]

An object proposal derived from one valid cached joint configuration.