Kyry10/microduck-arena-lab

Microduck Arena Lab

A small robot, three race courses, and a reproducible search for faster gaits.

We adapted a licensed running policy to the actual Microduck Arena, kept the policy interface intact, and selected checkpoints by exact five-seed evaluation. The roller entry reached official #1 at 1.648 s; its subsequent 1.516 s local candidate is included in this research snapshot. These are MuJoCo simulation results.

Watch the project film · Detailed workflow · Model downloads · Website kit

Results at the snapshot

Course	Best local mean	Finishes	Published Arena mean	Selected checkpoint
2 m sprint	1.084 s	5/5	1.096 s	D1 launchboost, 900
1.5 m hill, 8°	1.360 s	5/5	1.360 s	H1b adaptation, 500
2 m roller sprint	1.516 s	5/5	1.648 s, official #1	R2 polish, 1100

Local best and published entry can be different policies. The 1.516 s roller candidate was not yet submitted when this snapshot was assembled. Rankings are a dated snapshot, not a promise about the live leaderboard. See model metadata and recorded leaderboards for timestamps, per-seed times and checksums.

The sprint target was 0.988 s and was not reached in this snapshot. We include failed experiments because they explain the useful choices: preserving normalization, matching the race, measuring launch speed, and keeping nearby checkpoints instead of trusting the final training iteration.

Watch the actual policies

Film	What it shows
Project film	Three courses, real rollouts, and clearly labeled five-seed results
Sprint	Seed 5, 1.08 s timed finish; local mean 1.084 s
Hill	Seed 3, 1.34 s timed finish; local mean 1.360 s
Roller	Seed 3, 1.44 s timed finish; local mean 1.516 s

Clips include an untimed 0.50 s settle and labeled slow-motion replay. Video duration includes titles and replay, so it is not the race score. Footage comes from the Arena's unmodified physics loop; lighting, camera and visual-only meshes are presentation choices.

How we did it

1. Read the event, then measure baselines. Check course geometry, rig, start noise, actuator delays, command sweep, fall/timeout semantics and untimed settle.
2. Start from licensed weights. Hannes' microduck-running is Apache-2.0. Unlicensed reference policies were benchmarks only, never training parents.
3. Preserve the actor contract. Keep 61 observations, 14 actions and the exact normalizer. Imported ONNX actors get a fresh critic; our checkpoints resume the actor, critic and optimizer together.
4. Match the race. Start from STAND, use the real lane and slope/rollers, nominal BAM physics, narrow start/delay variation, forward reward, finish bonus and fall penalty.
5. Choose by the Arena. Score every 100 iterations, or 50 for polish. Only 5/5 policies can become an event best. Keep immutable selected artifacts.
6. Diagnose, then change one thing. We traced speed, tilt and wall contacts; lower-LR roller polish helped, while several static gait trims and later continuations did not.

Guides and code

• Complete workflow and reproduction
• Sprint experiment guide
• Hill experiment guide
• Roller experiment guide
• Experiment history and lessons
• Local workspace map
• Website video integration
• Authored experiment overlays and recorded results

This repository packages our experiment additions rather than another full copy of the upstream simulator. Install an overlay into a pinned upstream checkout to run the training scripts. Full weights live on Hugging Face; they are deliberately excluded from git.

Credits

Microduck and its simulation stack: Pollen Robotics. Licensed warm start: Hannes von Essen. Training: mjlab and rsl_rl. Evaluation: Microduck Arena.

Our experiment additions and guides are provided under Apache-2.0. Dependencies and robot assets retain their original licenses and notices; see NOTICE.