Sim-to-real transfer for legged locomotion at scale

A group at TU Berlin reports a new state of the art in sim-to-real transfer for quadruped locomotion, training policies entirely in simulation and deploying them zero-shot on hardware.

What's new

The team scaled domain randomization far beyond common practice, randomizing not just friction and mass but actuator latency, gear backlash, and even sensor timestamp jitter. Training ran across 4,096 parallel environments.

Key results:

• 93% success rate on unseen outdoor terrain, up from 71% for the previous best baseline
• Zero-shot transfer to two different quadruped platforms without retraining
• Recovery behaviors emerged without explicit reward shaping

Why it matters

Most Berlin robotics teams still budget weeks of on-robot fine-tuning per platform. If these numbers hold up in replication, that step could shrink dramatically.

"We stopped trying to make the simulator accurate and started making the policy indifferent to inaccuracy," the lead author writes.

Code and training configs are promised on GitHub after the conference deadline.