The robot in RoboDojo. This page is the robot manual ORA gives the model at the start of every episode, generated from ora/policy/embodiments/arx-x5/manual-public-geometry-v5.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 ARX X5: left 6-DoF arm, right 6-DoF arm; parallel-jaw grippers.
Controlled points
left_link6 is the left arm's controlled point, its wrist flange, not a fingertip; commands and measured poses already refer to it, so add no offset.
right_link6 is the right arm's controlled point, its wrist flange, not a fingertip; commands and measured poses already refer to it, so add no offset.
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_link6 local approach [1, 0, 0]; jaw-closing line [0, 1, 0]. Local Z is perpendicular to both. Current world directions are supplied with measured state.
right_link6 local approach [1, 0, 0]; jaw-closing line [0, 1, 0]. Local Z is perpendicular to both. 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 157.57 mm beyond the controlled point along +approach; the fingers begin 71.57 mm beyond it.
finger_1 collision bounds in that controlled-point frame are {"X":[71.57,157.57],"Y":[0.4016,37.7516],"Z":[-30.6401,30.3599]} mm; opening direction is [0.0, 1.0, 0.0]. These are bounds at the stated reference opening, not current contact surfaces.
finger_2 collision bounds in that controlled-point frame are {"X":[71.57,157.57],"Y":[-37.7556,-0.4056],"Z":[-30.6401,30.3599]} mm; opening direction is [0.0, -1.0, 0.0]. These are bounds at the stated reference opening, not current contact surfaces.
Nominal inner pad region in the controlled-point frame: 142.6–157.6 mm along the approach axis and -10.4–8 mm along local Z. The jaw contact centre is +0.1501 m along the approach axis from left_link6. Both are nominal source geometry, not a calibrated physical contact point.
Nominal full-open pad separation is 88.8 mm. Physical pad aperture remains uncalibrated. The public encoder reading is the mean of both finger joints on the normalized command scale (empty-close reference about 0.185, full-open 1); it is not measured pad separation, and a pair shifted to one side is no evidence of holding or release. A close directive is not a request for a measured zero, and reaching the encoder floor proves neither contact nor retention.
right_gripper: no single aperture is asserted across the tapered fingers.
At reference finger qpos [0, 0] m, the fingertips end 157.57 mm beyond the controlled point along +approach; the fingers begin 71.57 mm beyond it.
finger_1 collision bounds in that controlled-point frame are {"X":[71.57,157.57],"Y":[0.4016,37.7516],"Z":[-30.6401,30.3599]} mm; opening direction is [0.0, 1.0, 0.0]. These are bounds at the stated reference opening, not current contact surfaces.
finger_2 collision bounds in that controlled-point frame are {"X":[71.57,157.57],"Y":[-37.7556,-0.4056],"Z":[-30.6401,30.3599]} mm; opening direction is [0.0, -1.0, 0.0]. These are bounds at the stated reference opening, not current contact surfaces.
Nominal inner pad region in the controlled-point frame: 142.6–157.6 mm along the approach axis and -10.4–8 mm along local Z. The jaw contact centre is +0.1501 m along the approach axis from right_link6. Both are nominal source geometry, not a calibrated physical contact point.
Nominal full-open pad separation is 88.8 mm. Physical pad aperture remains uncalibrated. The public encoder reading is the mean of both finger joints on the normalized command scale (empty-close reference about 0.185, full-open 1); it is not measured pad separation, and a pair shifted to one side is no evidence of holding or release. A close directive is not a request for a measured zero, and reaching the encoder floor proves neither contact nor retention.
Sensors and cameras
head: fixed RGB-D; use calibration from the same observation.
left_wrist: arm-mounted moving RGB-D; use calibration from the same observation.
right_wrist: 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/robodojo/boundary-v3.md:
The exact task language supplied for the current episode is the objective. Current policy-visible scene evidence is authoritative; task names, layouts and prior episodes do not supply hidden geometry or a solution.
Alongside that language, this environment supplies a public completion contract for the current task: the official criteria restated in ordinary words. It says what the finished result must be and nothing about how to reach it; it is not a plan, an ordering, a hint or a solution. It adds no requirement the criteria do not carry and withholds none that they do, and it never stands in for the official verdict.
The scene is a simulated Isaac Sim tabletop in an episode-local world frame:
+X (red), +Y (green) and +Z (blue), +Z upward, positions in metres.
One 1.4 m by 1.1 m table stands with its top near z = 0.76 m (table body
centred at z = 0.74 m). The robot is mounted on the table's -Y side, so
+Y runs from the robot toward the far edge. The simulator's native w-first
quaternion order is never policy-visible; public orientations use the
[x,y,z,w] order stated by the embodiment card.
The public boundary exposes three RGB-D cameras with per-observation calibration and robot proprioception. Camera mounting, robot topology and command frames belong to the embodiment card.
Simulator object and actor state, ground-truth poses, segmentation, contacts, evaluator objects, reward internals, the success-predicate implementation, task source and the teleoperated source demonstrations are private and unavailable.
Only RoboDojo's reward manager supplies the official Boolean verdict, recorded by the Runtime at episode termination and not readable before it. Motion completion, closed jaws, a plausible arrangement and the policy lifecycle never replace it.
The episode ends by itself the moment the official verdict is true; nothing declares success. It also ends when one of the reported allowances is spent. It can end earlier, with the verdict false, when the task reaches a permanent official failure condition; which conditions a task carries and their thresholds are not policy-visible.
Where the task's completion contract requires it, both arms must be measured back at the poses they held when the episode reset, and they must be there at the same instant that everything else the task requires is also true. A finished arrangement left with an arm still out over the table is then not a success, and neither is a retraction taken before or after that finished state instead of together with it. Those reset poses are themselves the reference, and the allowance around them is not policy-visible. The advertised recovery action's home route is the deliberate way back.