Wojtek v2 locomotion
Simulation only. This policy runs in MuJoCo through the w01-tek
repository. It has no policy.npz pair, so the ROS stack cannot load it, and
the v6.27 legs are not on the real robot yet.
Run it from a fresh download:
python -c "from huggingface_hub import snapshot_download; \
snapshot_download('<org>/wojtek_v2-locomotion', local_dir='training/runs/wojtek_v2')"
mkdir -p training/runs/wojtek_v2/checkpoints
ln -s ../checkpoint training/runs/wojtek_v2/checkpoints/000367001600
./training/run.sh eval --run runs/wojtek_v2 --x-vel 0.5
./training/run.sh report --run runs/wojtek_v2
eval and report look for checkpoints/<step>/, and this repo stores the
checkpoint as checkpoint/, hence the symlink. They read the robot from
run.json and build the legs_v627 model themselves. Checked on
2026-09-22: a fresh download walks at 0.43 m/s on a 0.5 m/s command.
A joystick locomotion policy for Wojtek on the v6.27 four-bar legs (the
legs_v627 robot variant). It is the first keeper on these legs and the run
wojtek_legs_v627_loco_v2b from PR #31 of the w01-tek repository.
The legs are about 2.3 times longer than the stock ones. The robot stands 0.345 m tall and takes commands of -0.8 to 1.2 m/s forward, 0.5 m/s sideways, 1.0 rad/s of yaw and a stance height of 0.28 to 0.37 m.
What is here
checkpoint/: the Brax PPO checkpoint at 367M steps.run.jsonholds the resolved config it trained with.eval_report.json,eval_report.md: the fixed evaluation battery.videos/: the demo sequence, a 1.0 m/s trot and a 0.5 m/s arc at 0.6 rad/s.
There is no policy.npz and policy_meta.json pair. The policy observes the
gait clock, which the ROS runtime does not implement yet, so export
refuses it. The runtime also refuses a policy for these legs on the real
robot until the knee-clamp question is decided. This keeper runs in
simulation.
How it was trained
+experiment=legs_v627_locomotion, seed 1, 32768 envs, 367M steps from
scratch, on one RTX 5090 under MJWarp in about 18 minutes. On top of the
preset: reward.tracking_sigma=0.15, command.arc_prob=0.2,
command.pure_wz_prob=0.1, reward.scales.ang_vel_xy=-0.15 and
reward.scales.orientation=-10. The model stands level; the stance offset
that levels it is in robots.py at repository commit f30ddfc.
How it scores
No falls in the six battery scenarios.
| Measure | Value |
|---|---|
| Achieved speed at a 1.0 m/s command | 0.9 m/s |
| Velocity error, ramp / turn / arc | 0.055 / 0.16 / 0.28 m/s |
| Pitch and roll rocking on the ramp, standard deviation | 0.98 / 0.62 deg |
| Roll p95, turn / arc / stop | 2.0 / 1.8 / 5.1 deg |
| Torque p50 / p99 / max | 1.9 / 14.1 / 22.0 N*m |
| Front stride span | 0.208 m |
The yaw rate in the arc scenario is about 0.2 rad/s short of the 0.6 rad/s command. The stop rolls the body 5 degrees.