Files
UniversalTrain-ai/assets/wheeled_biped_example/README.md
T
chenlin c32df70326
Checks / Layout, translation and Colab checks (push) Has been cancelled
Checks / Ruff (push) Has been cancelled
Checks / Framework tests (native backend, CPU) (push) Has been cancelled
Checks / Portable Rust controller (push) Has been cancelled
release: v0.1 generic robot training and application platform
2026-11-05 11:05:01 +08:00

3.7 KiB
Raw Blame History

Wheeled biped example

This synthetic MuJoCo robot has two articulated legs, each with a pitch hip, pitch knee and continuous wheel hinge. The two tires provide the ground support. The battery case hangs below the upper frame and lowers the center of mass. At the calibrated neutral pose the 7.62 kg robot has a 0.421 m root height, a 0.19 m wheel radius and a 0.35 m track.

The locked wheeled-biped template applies position PD to the four leg joints and velocity feedback to the two wheels. The model declares 12 N·m hip, 16 N·m knee and 5 N·m wheel effort limits. Forward, reverse and yaw controls use the shared keyboard and gamepad vocabulary. The template's lateral command range is zero.

These commands reproduce the intake and stance checks. The last command writes a review copy of the generated task specification.

python scripts/onboard.py inspect assets/wheeled_biped_example \
  --profile assets/wheeled_biped_example/profile.json
python scripts/onboard.py calibrate --model assets/wheeled_biped_example/robot.xml \
  --config assets/wheeled_biped_example/stance.json --report out/wheeled_calibration.json
python scripts/onboard.py scaffold-tier3 --config assets/wheeled_biped_example/robot.json \
  --calibration out/wheeled_calibration.json --tasks-root out/wheeled_specs

In the two-second native stance check, both tires maintained support over the final measurement window, carrying 74.75 N of total normal load. The base settled 2.12 mm with no measured rotation, self-contact or engine warning. Calibration uses a bounded position hold for the wheels. Training and replay drive wheel velocity. In a separate hold using zero wheel velocity targets, an initial pitch of ±0.03 rad settled to within 0.00066 rad in two seconds. The peak effort per hip, knee and wheel was 0.291, 0.171 and 0.076 N·m respectively. Only the tires touched the ground.

Copy verify_policy.json beside the native export's controller.json and controls.yaml. Its forward, reverse, left yaw, right yaw and idle probes evaluate tracking after a one-second warmup. A policy export needs to pass those physical response checks before it can be packaged as a controlled app.

python scripts/train.py --task wheeled_biped_example --backend native --device cpu \
  --num-envs 4 --max-iterations 1 --headless --no-tensorboard
python scripts/export.py --task wheeled_biped_example --backend native --device cpu \
  --checkpoint <checkpoint.pt> --native-layout --out out/wheeled_policy
cp assets/wheeled_biped_example/verify_policy.json out/wheeled_policy/verify-input.json
python scripts/onboard.py verify --model out/wheeled_policy/robot.xml \
  --config out/wheeled_policy/verify-input.json --report out/wheeled_policy/verification.json

One training iteration exercises the pipeline. Use a trained checkpoint for the physical tracking probes.

A Warp run on an RTX 5080 used 1,024 environments and saved the selected checkpoint after 201 updates. With observation noise off, six-second Warp replays averaged +0.146 and −0.184 m/s for forward and reverse commands of ±0.18 m/s, and +0.308 and −0.347 rad/s for yaw commands of ±0.35 rad/s. All five replays, including idle, ran without resets. The exported actor passed the matching native verification probes.

The selected checkpoint exposes ±0.91 m/s forward and reverse and ±0.77 rad/s yaw. At full keyboard deflection, the native app averaged +0.866 and −0.927 m/s and +0.769 and −0.773 rad/s. The 0.19 m tires and 5 rad/s wheel target limit give a nominal rolling speed of 0.95 m/s. The final checkpoint after 1,000 updates averaged +0.899 m/s on a +1.30 m/s command. The published app uses the earlier checkpoint and its measured command range.