Robots and tasks

ALOHA-AgileX

The robot in RoboTwin. This page is the robot manual ORA gives the model at the start of every episode, generated from ora/policy/embodiments/aloha-agilex/manual-public-geometry-v3.md: what you read here is what the policy reads. Values measured during a run (poses, calibration, finger state) arrive with each observation instead; see Policy I/O and Frames and units.

Identity and topology

Fixed-base ALOHA-AgileX: left 6-DoF arm, right 6-DoF arm; parallel-jaw grippers.

Controlled points

left_tcp is the controlled reference, not a fingertip. Its local offset is [0.12,0,0] m from the declared manipulator flange; commands already target it. Do not add this offset again. right_tcp is the controlled reference, not a fingertip. Its local offset is [0.12,0,0] m from the declared manipulator flange; commands already target it. Do not add this offset again.

Frames and units

Positions are world metres; measured quaternions are [x,y,z,w]. Transform local geometry by the current measured pose; local lengths are not world-Z corrections.

End-effector axes

left_tcp local approach [1, 0, 0]; jaw-closing line [0, 1, 0]. Current world directions are supplied with measured state. right_tcp local approach [1, 0, 0]; jaw-closing line [0, 1, 0]. Current world directions are supplied with measured state. Local +approach points out of the gripper, from the palm toward the fingertips.

Motion capabilities

The ToolSurface owns motion, workspace, window and acceptance limits. Holding a command does not freeze measured state.

Gripper semantics

left_gripper: no single aperture is asserted across the tapered fingers. At reference finger qpos [0, 0] m, the fingertips end 35.57 mm beyond the controlled point along +approach, and the fingers reach back 50.43 mm from it toward the wrist. At that reference opening, the mesh centerline gap at TCP depth is 4.8964 mm; the gap varies with depth and is not a general minimum grasp width. right_gripper: no single aperture is asserted across the tapered fingers. At reference finger qpos [0, 0] m, the fingertips end 35.57 mm beyond the controlled point along +approach, and the fingers reach back 50.43 mm from it toward the wrist. At that reference opening, the mesh centerline gap at TCP depth is 4.8964 mm; the gap varies with depth and is not a general minimum grasp width.

Sensors and cameras

head_camera: fixed RGB-D; use calibration from the same observation. front_camera: fixed RGB-D; use calibration from the same observation. left_camera: arm-mounted moving RGB-D; use calibration from the same observation. right_camera: arm-mounted moving RGB-D; use calibration from the same observation.

Preservation and evidence limits

Robot geometry, command closure and encoder aperture establish neither contact nor retention. Endpoint arrival is not collision clearance or task success.

Dynamic facts and provenance

Current pose, finger measurements versus command targets, world axes, calibration, remaining budgets and execution thresholds come from current observation, Runtime or ToolSurface. Asset hashes and review/verification stay manifest-only.

What the policy is told about the environment

The environment boundary paired with this robot, from ora/policy/environments/robotwin/boundary-v4.md:

The exact task language supplied for the current episode is the objective. Current policy-visible scene evidence is authoritative; task names and prior layouts do not provide hidden geometry or a solution.

The public boundary exposes selected rendered RGB-D sensors, calibration and robot proprioception. Camera mounting, robot topology and command frames belong to the embodiment card.

Simulator object state, actor identities, segmentation, contacts, task source, expert programs, demonstrations and the success-predicate implementation are private and unavailable.

Only the environment evaluator at episode termination supplies the official Boolean verdict. Motion completion and the policy lifecycle never replace it. An in-episode completion aid, where one is advertised, is advisory and does not bind that verdict in either direction.

What this boundary does not guarantee

Measured robot and scene state can change without a command from you. Contact, gravity and settling all act between one observation and the next, so a changed measurement does not prove a command took effect and an unchanged one does not prove that nothing happened.

A rendered view reports what that sensor sees from where it is. An object missing from one view may be outside its field, behind another object, or past the usable depth range. Missing from a view is not missing from the scene, and this boundary exposes no channel that could confirm a scene defect; change viewpoint rather than reasoning about whether the rendering is intact.

Nothing here announces that the objective has become unreachable. The absence of such a signal is not evidence that the episode is still recoverable.