Acrobot swing-up

dmcontrol-acrobot-swingupdmcontrolDeepMind Control Suite bodiesManipulationhard

Reference solution, 1,065 of 3,000 steps.

Instruction

An acrobot from the DeepMind Control Suite (MuJoCo): a two-link pendulum in the x-z plane (z up). The upper arm hangs from an unactuated shoulder hinge at (x, z) = (0, 2); the lower arm is attached at the elbow, the only motor. Each link is 1 m long and weighs 1 kg; both joints have light damping. Angles are in radians: shoulder_angle 0 = upper arm pointing straight up (+-pi = hanging down), elbow_angle is the lower arm's angle relative to the upper arm (0 = straight); positive angles turn towards +x. The target is the point straight above the shoulder at full height, (0, 4). 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 acrobot up until the tip of the lower arm reaches the target at the top and keep it there. It starts hanging nearly straight down (shoulder_angle = -3.08, elbow_angle = 0.01 rad), at rest.

Success: the tip of the lower arm inside the target sphere (tip_to_target <= 0.2 m, dm_control's swingup_sparse condition), held for 20 consecutive steps (0.2 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 TAU [--repeat N] applies a torque of 2 N m per unit at the elbow, TAU in [-1, 1] (positive increases elbow_angle), for N steps of 10 ms. The motor is weak: the acrobot has to be pumped up by swinging, and balancing it upright needs fast, precise feedback. There are no skills.

Observation. robo observe reports dm_control's own observation, orientations = [sin shoulder, sin(shoulder + elbow), cos shoulder, cos(shoulder + elbow)] and velocity = [shoulder, elbow angular velocity (rad/s)], plus shoulder_angle, elbow_angle, shoulder_vel, elbow_vel, the tip of the lower arm tip_pos and target_pos as (x, z) in metres, target_radius and tip_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, lookat; robo observe --image [--camera C] saves a picture.

The step budget is 3000 steps (30 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-acrobot-swingup

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

No trials yet.