Ball in cup

dmcontrol-ball-in-cup-catchdmcontrolDeepMind Control Suite bodiesManipulationhard

Reference solution, 355 of 1,000 steps.

Instruction

Ball in cup from the DeepMind Control Suite (MuJoCo), in the x-z plane (z up, gravity on). A cup (0.065 kg; an open V-bottomed cup, about 8 cm wide inside the rim and 10 cm deep, opening upwards) is held by springs: it slides in x and z about its home position (0, 0.6) with a stiffness of 20 N/m and damping of 3 N s/m per axis. A ball (0.065 kg, 2.5 cm radius) hangs from the bottom of the cup on a 0.3 m string, which goes slack when the ball comes closer. The cup's position cup_pos is the centre of its rim (the bottom is 0.1 m lower). The ball starts at rest somewhere below the cup. Simulated time advances only when you act; the scene is paused while you think. A feedback controller can run as a script: Python 3 with numpy is available, and robo observe --json / robo act ... --json print machine-readable output.

Task

Swing the ball on its string into the cup and keep it there. The ball starts at (-0.084, 0.553), at rest, and falls when the episode starts; the cup is at its home (0.000, 0.600).

Success: the ball entirely inside the cup: its centre within 0.025 m of cup_target_center along both x and z (dm_control's own in-target check: the target box's half-size 0.05 m minus the ball radius), held for 50 consecutive steps (1 s of simulated time), judged by the episode server from the simulated state after every step. The episode ends as solved the moment that happens; robo done before that scores 0.

Controls. robo act FX FZ [--repeat N] pushes the cup with 5 N per unit along x and along z, each in [-1, 1], for N steps of 20 ms. The springs limit how far the cup can go (about 0.25 m from home at full force). There are no skills.

Observation. robo observe reports dm_control's own observation, position = [cup x, cup z, ball x, ball z] (joint positions: cup relative to its home, ball relative to its reference 0.2 m above the floor) and velocity, plus by name (world frame): cup_pos, cup_vel, ball_pos, ball_vel, cup_target_center (the middle of the inside of the cup) and string_length. Every task also reports reward (dm_control's own reward for the last step, for information), in_target (whether the success condition holds right now), hold_steps (for how many consecutive steps it has held) and hold_required, and time_s (simulated time). Cameras: cam0, cam1; robo observe --image [--camera C] saves a picture.

The step budget is 1000 steps (20 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.
The episode ends as solved the moment the task's success rule holds (see Success); robo done before that scores 0.
Call robo done exactly once when finished.

Run this task

bench eval run \
  -d google-deepmind/dm-control@0.1 \
  --registry https://robouse.ai/hub/registry.json \
  --agent oracle \
  --include dmcontrol-ball-in-cup-catch

Pinned to robohub commit bcd0d002b9df. The verifier and the reference solution are not published.

No trials yet.