Robo Use

One robot holds a peg, the other a plate with a hole; get the peg through the hole

robosuite-twoarm-peginhole-pandarobosuiteFranka PandaManipulationhard

Reference solution, 120 of 500 steps.

Instruction

Two Franka Emika Panda (7-DoF) with its parallel-jaw gripper arms face each other across a table: robot0 on the -y side, robot1 on the +y side. The scene is robosuite's TwoArmPegInHole environment (MuJoCo).

Goal: Robot0 holds a peg (a cylinder fixed to its hand) and robot1 holds a square plate with a hole (fixed to its hand). Bring them together so the peg passes through the hole. robosuite's TwoArmPegInHole check: the peg's axis is nearly parallel to the hole's axis (|cos| > 0.95), the peg's centre is within 6 cm of the hole's axis, and it sits between 12 cm in front of and 14 cm behind the hole. Both must still hold after robo done.

Call robo done when finished; the check is made after the robot then holds still for 10 steps, so the result must last.

This robot

  • The action has 12 numbers: robot0's DX DY DZ DROLL DPITCH DYAW, then robot1's, each in [-1, 1]: robosuite's OSC_POSE end-effector commands, 20 steps per second, all in the world frame. DX/DY/DZ = 1.0 asks for a 5 cm move (held at 1.0 a hand travels about 1.1 cm per step); DROLL/DPITCH/DYAW rotate that hand about the world x/y/z axes (1.0 asks for 0.5 rad).
  • robo move-to and robo grip are not available (two robots); use robo act.
  • World frame, metres: +x points away from the table edge between the robots, +y to the robot1 side, +z up.
  • robo observe fields: for each robot i in 0, 1: roboti_hand_pos, roboti_hand_quat (x, y, z, w), and, if it has a gripper, roboti_gripper_open (finger-pad distance, metres) and roboti_gripper_yaw_deg (direction of the fingertip line in the table plane, degrees from +x, in [-90, 90)) (neither robot has a gripper); peg_pos (centre of the peg), peg_axis (unit vector along the peg), peg_quat, hole_pos, hole_quat, hole_center (centre of the hole), hole_axis (unit normal of the plate through the hole), and robosuite's own alignment numbers: peg_hole_cos (|cos| of the angle between the two axes), peg_hole_d (distance of the peg's centre from the hole's axis, m), peg_hole_t (position of the hole's centre along the peg's axis relative to the peg's centre, m).
  • robo observe --image saves the frontview camera; add --camera robot0_eye_in_hand for the wrist camera. Images are 320x320.
  • The step budget is 500 steps.
  • The action has 12 numbers: 6 for robot0, then 6 for robot1 (no grippers). At the start the peg points down and the plate stands upright, so the axes are about 90 degrees apart; turning one wrist the whole way is hard, so share the turn between the two arms.
How the robot is controlled and scored

You are controlling a simulated robot. Read the task below, then solve it by running the robo command in your shell (start with robo info and robo observe). Keep going until the task is done, then call robo done once. Do not stop to ask questions; there is no human to answer.

How to control the robot

You are the robot's policy. You act only through the robo command in your shell. There is no other way to move the robot, and you cannot read or change the simulator, the scoring, or other files to succeed; the episode server judges the final physical state itself.

robo info                         # action space, available skills, step budget
robo observe                      # robot and object state as numbers
robo observe --image              # also saves a camera image and prints its path (open it to look)
robo act DX DY DZ GRIP [--repeat N]   # low-level action, applied N times (N <= 50)
robo move-to X Y Z [--grip G]     # skill: move the gripper toward a point (if enabled for this task)
robo grip G [--steps N]           # skill: hold position and set the gripper (+1 close, -1 open)
robo done "short summary"         # end the episode and ask for scoring
robo give-up "reason"             # end the episode without claiming success
  • Positions are in metres in the world frame (x, y on the table plane, z up).
  • The episode has a fixed step budget (see robo info); every simulated step counts, including skills.
  • Unless the task says otherwise, success is judged about 10 steps after you call robo done, with the robot holding still, so the goal must still be true when the robot stops.
  • Work in small steps and re-observe after each motion. Call robo done exactly once when finished.

Run this task

ROBOUSE_ORACLE_TOKEN=$(openssl rand -hex 16) \
  bench eval run \
    -d arise-initiative/robosuite@0.2 \
    --registry https://robouse.ai/hub/registry.json \
    --agent oracle \
    --include robosuite-twoarm-peginhole-panda

Pinned to robohub commit e472b1a1e041. The verifier and the reference solution are not published.

Trial
GPT-6 Astra · CodexSolved149 / 5002 minTrial Compare
Kimi K3 · Claude CodeSolved105 / 5005 minTrial Compare
Kimi K3 · mini-swe-agentSolved453 / 50012 minTrial Compare
GLM-5.3 · mini-swe-agentSolved203 / 50022 minTrial Compare
GLM-5.3 · Claude CodeSolved295 / 50012 minTrial Compare
GLM-5.3 · CodexInfra failure132 / 5002 minTrial Compare

These trials ran on version 0.1, now deprecated (superseded by 0.2).