Reacher (small target)

dmcontrol-reacher-harddmcontrolDeepMind Control Suite bodiesManipulationeasy

Reference solution, 55 of 500 steps.

Instruction

A planar two-link arm from the DeepMind Control Suite (MuJoCo) lies flat on a table (the x-y plane, z up), inside a square arena with walls at x, y = +-0.3 m. The shoulder is at the origin and turns without limit; the upper link is 0.12 m long, the wrist (range +-160 degrees) carries a 0.12 m link ending in the fingertip (a 1 cm sphere). Angles are in radians, counter-clockwise seen from above: shoulder_angle is the upper link's direction from +x, wrist_angle the lower link's angle relative to the upper link. The links are light, so the torques are small. 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

Move the fingertip into the target sphere at (-0.079, -0.105) and keep it there. The arm starts at rest with the fingertip at (0.122, 0.156), 0.330 m from the target; target_radius = 0.025 m.

Success: the fingertip within target_radius of the target's centre (fingertip_to_target <= target_radius, the task's own reward condition), held for 25 consecutive steps (0.5 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 SHOULDER WRIST [--repeat N] applies torques of 0.05 N m per unit at the shoulder and the wrist, each in [-1, 1] (positive = counter-clockwise), for N steps of 20 ms. There are no skills.

Observation. robo observe reports dm_control's own observation, position = [shoulder, wrist angle] (rad), to_target = target minus fingertip (x, y in m) and velocity = [shoulder, wrist angular velocity] (rad/s), plus shoulder_angle, wrist_angle, shoulder_vel, wrist_vel, fingertip_pos, target_pos (x, y in m), target_radius (the target sphere's radius plus the fingertip's) and fingertip_to_target (m). 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: fixed; robo observe --image [--camera C] saves a picture.

The step budget is 500 steps (10 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

ROBOUSE_ORACLE_TOKEN=$(openssl rand -hex 16) \
  bench eval run \
    -d google-deepmind/dm-control@0.2 \
    --registry https://robouse.ai/hub/registry.json \
    --agent oracle \
    --include dmcontrol-reacher-hard

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

No trials yet.