Testing different training methods for four-legged robots

Fourteen methods were compared using five different quadrupeds on various 3D terrains.

Fourteen methods were compared under one protocol. These methods include unsupervised environment design (UED), domain-randomisation, and robotics curricula.

The study used five 3D quadrupeds in MuJoCo MJX. These robots are the closed-chain W01-TEK and the open-chain Unitree Go1, Unitree Go2, Google Barkour vB, and Boston Dynamics Spot.

Levels consist of vectors of eight interpretable terrain parameters. All curricula drive the same PPO student at matched environment-step budgets.

Methods with easy or low-difficulty training levels often fill the top groups on in-distribution random levels. This was true for Go1, Barkour, and W01-TEK, except for PAIRED+BiBC.

On three robots, the leader on in-distribution random levels is not the leader on held-out terrain. No method is in the top group on every robot and level set.

Key facts

  • Fourteen UED, domain-randomisation, and robotics curricula were compared.
  • The evaluation used five 3D quadrupeds including Unitree Go1, Unitree Go2, Google Barkour vB, and Boston Dynamics Spot.
  • The leader on in-distribution random levels was not the leader on held-out terrain for three robots.

All sections

Edition