Stack blocks in a stated order: ALOHA 2 in the tabletop
stack-in-order--aloha-2--tabletop--s4seed 4 variantremixALOHA 2Manipulationmedium
Instruction
You control a ALOHA 2 (two ViperX 300 s arms) (mounted on the table (top at z = 0.75 m), facing south: the left arm at (-0.4, 0.2); the right arm at (0.4, 0.2)). Two ViperX 300 s arms side by side on a work surface; each has its own gripper target and gripper. Together they span more than either arm reaches alone. The scene is a 2 m x 0.8 m table in a plain room; arms mount at its rear edge, floor robots reach it from the south, drones take off from a pad in the corner. Everything is simulated with MuJoCo physics: objects are free rigid bodies that move only through contact and friction, so a loose grip or a collision can drop or knock things over. World frame in metres: +x east, +y north, +z up, floor at z = 0 (the floor area spans x -2.5..2.5, y -2.2..1.6); headings in degrees counter-clockwise from +x.
Task
Build a tower of the three blocks on the black mark: the blue block at the bottom, the red block on top, and the remaining block in the middle.
Success: the bottom block rests on the surface with its centre on the black mark (within 3 cm), each other block rests on the one below it (centres within 2.5 cm horizontally, touching), every block is level (tilted less than 15 degrees) and the robot touches none of them; and no safety event happened (a hard collision with furniture or walls, an object falling to the floor, the robot falling over, or harm to a person; safety_events in robo observe lists them). Judged by the episode server from the simulated state after you call robo done and the robot has held still for 10 steps.
Controls. robo act L_DX L_DY L_DZ L_G R_DX R_DY R_DZ R_G (8 values; one step is 50 ms; --repeat N holds it for N steps). left.ee_delta L_DX, L_DY, L_DZ (each in [-1, 1]; x 2 cm per step): moves the gripper target (world axes); IK follows it with the gripper pointing down. left.gripper L_G (each in [-1, 1]; command): 0 keeps the fingers, > 0 closes, -f opens to fraction f (-1 fully open). right.ee_delta R_DX, R_DY, R_DZ (each in [-1, 1]; x 2 cm per step): moves the gripper target (world axes); IK follows it with the gripper pointing down. right.gripper R_G (each in [-1, 1]; command): 0 keeps the fingers, > 0 closes, -f opens to fraction f (-1 fully open). All zero holds still. Skills (robo skill NAME ARG ...; robo info lists them with their arguments): move_to X Y Z [ARM=auto]: move the grasp point (between the fingertips) to (X, Y, Z) along a straight line, gripper pointing down; reports whether it got there (it stops if blocked or out of reach); grasp [ARM=auto]: close the fingers until they stop; reports what is held; release [ARM=auto]: open the fingers fully; home [ARM=auto]: move the gripper back to its start point above the work area.
Observation. robo observe reports robot (what the robot is, its capabilities and grasp, its workspace, and its current state: gripper points, targets and openings per arm), objects (per task object: kind, colour, centre pos, tilt, yaw, size, mass and resting_on: a surface, a fixture, another object, gripper or the floor), fixtures (containers with their inner size and rim height, pegs, marks, pads, tags), scene (surfaces with their top heights and sizes, obstacles, the floor area) and safety_events. Cameras (robo observe --image --camera NAME): overview, robot/left_wrist, robot/right_wrist, front, top.
The step budget is 1400 steps (70 s of simulated time).
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 # the robot, its sensors, action groups, skills and step budget
robo observe # robot and scene state as numbers
robo observe --image [--camera C] # also saves a camera image and prints its path (open it to look)
robo act V1 V2 ... [--repeat N] # one low-level action (the action groups under Controls), applied N times (N <= 50)
robo skill NAME ARG ... # run a skill listed by `robo info`; it runs until it finishes and reports the result
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 (+z up); angles are in degrees unless a field says otherwise.
- The episode has a fixed step budget (see
robo info); every simulated control step counts, including the steps a skill runs. - Skills are ordinary controllers: they can fail, stop early or be blocked by the scene. Read what they report and re-observe.
- Success is judged about 10 steps after you call
robo done, with the robot holding still (each action group's hold value: zero for velocity and delta commands, full brake for a car), so the goal must still be true when the robot stops. - Call
robo doneexactly once when finished.
Run this task
bench eval run \
-d benchflow/robouse-composed@0.1 \
--registry https://robouse.ai/hub/registry.json \
--agent oracle \
--include stack-in-order--aloha-2--tabletop--s4Pinned to robohub commit 1f901b956f61. The verifier and the reference solution are not published.