Skip to main content

Example: Wave Arms

Before you start, complete the station setup.

Run the bundled motion policy​

Try a complete evaluation without supplying a model. On a bimanual YAM station, the bundled straight-line policy sweeps the left arm along base-frame Y, from −20 cm to +20 cm and back, five times. The right arm stays at zero joint angles, both grippers stay open, and the moving tool's orientation stays fixed. Confirm the sweep and starting pose with the operator.

se3labs eval run \
  --station example-station-id --address example-station-addr \
  --request se3labs.sdk.examples.straight_line:STRAIGHT_LINE_REQUEST \
  --output straight-line-results.json

Each cycle takes ten seconds. The episode budget is 52 seconds, including a final hold. The SDK includes MuJoCo, SciPy, and the YAM kinematic model; the policy computes the trajectory before rollout and sends it on its first inference. The operator then scores whether it completed all five cycles.

Move both arms through the harness​

With ACT Harness serving a live evaluation, this sequence moves each arm 5 cm left of its starting position, then 5 cm right, then back to the start. Commands run in order, one arm at a time, and stop at the first failure. In each arm's base frame, positive Y is left and negative Y is right.

python -m se3labs.sdk.examples.act_harness wait
python -m se3labs.sdk.examples.act_harness run \
  "goto left 0 0.05 0" \
  "goto left 0 -0.05 0" \
  "goto left 0 0 0" \
  "goto right 0 0.05 0" \
  "goto right 0 -0.05 0" \
  "goto right 0 0 0"
python -m se3labs.sdk.examples.act_harness end-episode

You can give this same sequence to an agent as its evaluation task. goto coordinates are offsets from the episode's starting pose. end-episode requests the station's home walk after scoring.

Continue with reading the evaluation results.