Push the puck to the goal, around the wall between them

metaworld-push-wall-visionvision variantVisionSawyer armManipulationVisionmedium

Reference solution, 81 of 500 steps.

Instruction

Vision variant

This is the vision version of the task. robo observe does not give object or goal positions: you get only the robot's own state (hand_pos, gripper_open) and two camera images (corner, topview), saved on every robo observe and printed as paths. Open the images to see the scene. robo info lists each camera with a 3x4 projection matrix P: for a world point (x, y, z), [u*w, v*w, w] = P @ [x, y, z, 1] gives its pixel (u, v) in that camera's saved image (u to the right, v down, origin at the top-left). With two cameras you can triangulate a point you see in both, or intersect a pixel's ray with a known height. Any field names in the description below that are not in your observation are hidden in this variant.

A Sawyer robot arm with a two-finger gripper stands at a table (adapted from Meta-World, push-wall-v3).

Goal: Push the puck to the goal, around the wall between them.

In robo observe: hand_pos is the gripper position, gripper_open its opening (1 = open), obj1_pos/obj1_quat and obj2_pos are the task objects (obj1 = puck; goal = target), and goal_pos is the target. The full 39-number observation vector is also given as obs_vector.

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 benchflow/robouse-tabletop@0.2 \
    --registry https://robouse.ai/hub/registry.json \
    --agent oracle \
    --include metaworld-push-wall-vision

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

Filter models

					Trial
GPT-6 Astra · Codex	Not solved	500 / 500	3 min	Sep 27, 2026	Trial Compare
Kimi K3 · Claude Code	Not solved	500 / 500	12 min	Sep 28, 2026	Trial Compare
Kimi K3 · mini-swe-agent	Not solved	500 / 500	13 min	Sep 28, 2026	Trial Compare
GLM-5.3 · mini-swe-agent	Not solved	382 / 500	10 min	Sep 27, 2026	Trial Compare
GLM-5.3 · Claude Code	Not solved	25 / 500	15 min	Sep 27, 2026	Trial Compare

These trials ran on version 0.1, now deprecated (superseded by 0.2).