Crown Roll
I wanted my Microduck to skate into a forward roll, go right over its head and land straight back on its wheels. The whole act takes 3.48 seconds. It runs on the roller rig in the Got a Talent event.
In the filmed run, both skates touch down together, 45 milliseconds after the head leaves the floor. The body clears the ground and the duck skates away at about 0.50 m/s. Watch the side view.
How I trained it
I used my Roller Sprint checkpoint 500 for the run-up and skate-away. I first trained the roll itself by trying hip, knee, ankle and head movements in the Arena simulator. Rewarding forward speed on its own let the duck scrape along on its back. I started checking what actually touched the floor after it rolled over its head.
I kept the motor sequence from my favourite run and fitted a first pass at the feedback from 128 other simulated starts. I then trained 12 feedback gains by episode reward, using cross-entropy search. The policy reads tilt, angular velocity and joint positions to adjust the motors during the roll. The timeline supplies the timing.
On 64 new simulated starts, the exported policy landed straight on its skates and rolled away on 62. The same motor sequence with the roll feedback switched off did it on 58 of those starts. The model reproduced all 174 frames of the filmed run in my local Arena. I tested the straight approach in simulation. Individual results.
Train or export it
I used Python 3.12, Microduck Arena 0.7.0, ONNX 1.23.0 and ONNX Runtime 1.24.4. From this model folder in the Arena environment:
python training/train.py --output my-checkpoint.json --workers 12
python training/export.py my-checkpoint.json my-policy
To rebuild the submitted policy from the saved gains:
python training/export.py checkpoint.json rebuilt
The training folder includes the roll reference and initial feedback. The scripts fetch the pinned sprint policy and check its checksum. The export produces the ONNX and the timeline. Training code · Selected checkpoint · Pollen licence
- Downloads last month
- 15