Robots and tasks

Frames and units

For each environment ORA drives: the world frame, the point a pose command moves, the gripper axes, quaternion order, units, cameras, depth and control timing. The values match the robot manuals the model reads (for example ARX X5) and the default tool surfaces. A value marked not published is one ORA does not give the policy. For the message shapes that carry these values, see Policy I/O schema. For installation, see the environment pages for LIBERO, RoboTwin and RoboDojo.

Conventions shared by every environment

Quantity Convention
Positions World frame, metres
Orientation Quaternion [x, y, z, w]; normalized on input (norm 0.98–1.02 accepted, then renormalized)
Measured pose pose_xyz_xyzw: 7 numbers, position then quaternion
Angles Radians in canonical records, receipts and controllers. On the three default robot surfaces (policy_angles_degrees_v1) the copy the model reads states every angle in degrees. Inputs named rotation_angle_deg / angle_deg are degrees; rotation_rotvec_rad / rotation_vector_rad stay radians
Pixels OpenCV, origin top-left, u right, v down, native RGB resolution
Camera frame OpenCV: x right, y down, z forward
Intrinsics intrinsic_cv, 3x3, maps camera coordinates to native pixels
Extrinsics extrinsic_cv, 4x4, world-to-camera; cam2world_cv is its inverse, computed in plain Python so its digits do not depend on the host's BLAS
Depth float32 array, metres, optical-axis z distance (not ray length); delivered on demand to tools, not inlined each step
Calibration delivery Per observation, for every camera, fixed or moving

Each controlled point also has two local axes: approach (the direction the fingers point) and jaw closing line. The measured state carries their current world directions (axes.approach_world, axes.closing_world), so a policy does not need to rotate local axes itself.

LIBERO / LIBERO-Pro (Franka Panda)

Item Value
Robot Fixed-base Franka Panda (MountedPanda, PandaGripper), one 7-DoF arm, public arm id right
World frame "LIBERO world": right-handed, +Z up, metres. Origin location and X/Y directions: not published
Robot base pose Owned by the runtime and used by the planner; not published to the policy
Action envelope x [-0.60, 0.60], y [-0.50, 0.50], z [-0.10, 1.30] m (rejection bound, not reach)
Controlled point right_grip_site (robosuite grip_site), the grip centre, 0.097 m along local +Z from the end-effector root. Commands already target it
Local axes Approach [0, 0, 1] (local +Z, out of the gripper); jaw closing line [1, 0, 0] (local +X). Top-down means local +Z toward world -Z
Gripper command Named directive open / close on the default surface
Gripper measurement Two finger displacements (m); measured_aperture_mm = sum of their magnitudes x 1000; measured_opening_normalized = mean magnitude / 0.04, clipped to [0, 1], 0 closed, 1 open. Commanded opening is shown beside it
Maximum aperture 0.08 m jaw separation
Cameras agentview: fixed RGB-D. wrist (source robot0_eye_in_hand): mounted on the gripper, moves with it. Main camera agentview
Image size 256 x 256 (rendered rows are flipped to top-left order before publication)
Depth Metric depth converted from the renderer's z-buffer; valid_depth_range_m (near, far) is recorded per camera
Control rate One control step = 25 physics steps of 2 ms = 50 ms (control_freq 20)
Per command wait_ticks: 1–20 control steps (up to 1.0 s). Planned motion: up to 4000 execution ticks and 2000 waypoints
Arrival tolerance 0.005 m, 0.05 rad

RoboTwin (ALOHA-AgileX)

Item Value
Robot Fixed-base dual-arm ALOHA-AgileX, two 6-DoF arms left, right
World frame World metres. Origin and axis directions: not published to the policy. ORA's own task reviews, which the policy does not see, take world +Z as up and negative Y as nearer the robot
Action envelope x [-0.6, 0.6], y [-0.4, 0.4], z [0.7, 1.45] m
Controlled point left_tcp / right_tcp: grip-centre TCP, 0.12 m along local +X from the link6 flange. Commands already target it
Local axes Approach [1, 0, 0] (local +X); jaw closing line [0, 1, 0] (local +Y)
Native quaternion The simulator's [w, x, y, z] is converted to [x, y, z, w] before publication
Gripper command open / close per arm; internally open maps to 1, close to 0
Gripper measurement gripper_opening_m: measured jaw separation in metres; gripper_state uses a normalized 0 (closed) to 1 (open) scale that is not an aperture. No single maximum aperture is asserted for the tapered fingers
Cameras head_camera, front_camera: fixed RGB-D. left_camera, right_camera: mounted on the grippers, move with them. Main camera head_camera
Image size Set by the upstream task configuration; not published in ORA
Depth Simulator depth in millimetres, divided by 1000 and stored as metres
Control rate Physics step 4 ms; one tick = 10 physics steps = 40 ms. A policy-level control rate is not published
Per command wait_ticks: 1–20 ticks (up to 0.8 s). Planned motion: up to 20000 physics steps (global) or 1280 (local)
Arrival tolerance 0.005 m, 0.05 rad, checked at the TCP

RoboDojo (ARX X5)

Item Value
Robot Two fixed-base ARX X5 arms, 6-DoF each, left, right
World frame Episode-local Isaac Sim frame, metres; +X red, +Y green, +Z blue and up. Robot on the table's -Y side, so +Y runs from the robot to the far edge
Scene Table 1.4 m x 1.1 m, top near z = 0.76 m (0.765 m in the workspace profile), spanning x [-0.7, 0.7], y [-0.6, 0.5]. Arm bases at (±0.3, −0.45, 0.765); link6 reset poses at (±0.3, −0.352, 0.922); head camera at z = 1.308 m
Action envelope x [-0.75, 0.75], y [-0.65, 0.55], z [0.72, 1.40] m
Controlled point left_link6 / right_link6: the wrist flange, offset 0. Not the fingertips: the nominal jaw contact centre is 0.1501 m along the approach axis
Local axes Approach [1, 0, 0]; jaw closing line [0, 1, 0]; local Z perpendicular to both
Native quaternion Simulator order is [w, x, y, z] and never policy-visible
Gripper command Range [0, 1], 0 closed, 1 open; default surface uses open / close. An explicit command moves the drive target at most 0.2 (normalized) from the current encoder reading
Gripper measurement measured_opening_normalized: mean of both finger joints on the command scale (empty close about 0.185, full open 1). measured_aperture_mm is null: physical pad aperture is uncalibrated. Nominal full-open pad separation 88.8 mm
Cameras head (source cam_head): fixed RGB-D. left_wrist, right_wrist: mounted on the arms, move with them. Main camera head
Image size Not published in ORA
Depth Metres, distance_to_image_plane; finite_depth_fraction recorded per camera
Control rate Physics 4 ms; 25 Hz control; one native action = 10 physics steps
Per command A move may use up to 250 native actions of the task's native-action allowance, other commands 20 or 30; the receipt reports how many were used, a refused command uses none
Arrival tolerance 0.01 m, 0.10 rad, gripper 0.03 (normalized); published each step in controller_feedback

Synthetic

Synthetic has no physics and no official verdict (official_success: null). It checks action submission and observation sequencing.

Item Value
World frame World metres, quaternion [x, y, z, w]; origin and axes not published
End effectors Observation carries left and right, each starting at pose [0, 0, 1, 0, 0, 0, 1] with gripper 1.0. The target-state manual treats right as the single controllable grip site
Action envelope Quickstart: x [-0.6, 0.6], y [-0.4, 0.4], z [0.7, 1.45] m
Gripper Number in [0, 1], 1 open, 0 closed
Cameras runtime.policy_cameras, default one fixed head_camera. The target-state manual describes fixed agentview and moving wrist
Images and depth 4 x 3 RGB test pattern; constant depth 1.0 m; identity intrinsics and extrinsics
Control rate Not published (no simulated time). The quickstart ends after one accepted action

Cross-environment comparison

LIBERO / LIBERO-Pro RoboTwin RoboDojo Synthetic
Arms 1 (right) 2 2 2 in state, 1 controllable
Controlled point grip site (grip centre) TCP, flange + 0.12 m link6 flange grip site
Approach / jaw line +Z / +X +X / +Y +X / +Y not published
World up +Z +Z (maintainer notes only) +Z not published
Quaternion (policy) xyzw xyzw xyzw xyzw
Gripper measurement aperture mm and normalized opening jaw separation m normalized encoder, no aperture normalized value
Main camera agentview head_camera head head_camera
Wrist cameras wrist left_camera, right_camera left_wrist, right_wrist none by default
Native depth source z-buffer to metres mm to metres metres constant
Time unit 50 ms control step 40 ms tick 40 ms native action none
Arrival tolerance 5 mm / 0.05 rad 5 mm / 0.05 rad 10 mm / 0.10 rad not published

Silent mistakes

Differences that break code written for one environment

  • The controlled point is not the same physical point. A RoboDojo pose moves link6, about 150 mm behind the jaw contact centre. A RoboTwin pose moves a TCP already 0.12 m ahead of the flange; LIBERO moves the grip centre. Code that adds a fingertip offset on every robot double-counts it on two of them.
  • Approach axes differ. Panda approaches along local +Z; ALOHA and ARX approach along local +X. A "top-down" quaternion copied from LIBERO points the RoboTwin or RoboDojo gripper sideways. Use the published approach_world and closing_world.
  • Gripper numbers are not apertures by default. Only LIBERO (measured_aperture_mm) and RoboTwin (gripper_opening_m) publish a measured separation, in different units. RoboDojo's 0.185 is an encoder floor, not 0 mm.
  • Internal gripper scales run in opposite directions. The measured opening is 0 closed / 1 open everywhere it is published, but the Panda motion profile writes open = 0.0 and close = 1.0. Send the named open / close directives, not numbers.
  • Angle units change between what you send and what you read. Default surfaces read back degrees, accept rotation_angle_deg in degrees, and keep rotation_rotvec_rad in radians.
  • extrinsic_cv is world-to-camera. Back-projection needs cam2world_cv. RoboTwin and Synthetic calibration files also contain cam2world_gl (OpenGL axes); do not mix it with OpenCV pixels.
  • A "tick" is not one duration. 50 ms on LIBERO, 40 ms on RoboTwin and RoboDojo, and on RoboDojo each tick also spends the task's native-action allowance.
  • Workspace envelopes differ in height. LIBERO z starts at −0.10 m; RoboTwin and RoboDojo start near the table at 0.70–0.72 m.