Towards bridging the gap: Systematic sim-to-real transfer for diverse legged robots

This paper introduces PACE, a systematic sim-to-real pipeline for legged robots that builds accurate models of real hardware from the bottom up. The approach begins with detailed actuator characterization — including delay, torque dynamics, and friction — followed by robot-level parameter identification using evolutionary optimization. With this calibrated simulation, we train reinforcement-learning locomotion controllers that are fully aware of real actuator limits and dynamic behavior.

We demonstrate that this pipeline transfers directly to multiple quadruped platforms, achieving robust, energy-efficient locomotion on real hardware without fine-tuning. Across robots with different sizes, masses, and actuator designs, PACE consistently reduces tracking errors, improves stability, and lowers energy consumption compared to standard simulation baselines. The results show that accurate actuator modeling is a key missing ingredient for reliable, scalable sim-to-real transfer in legged robotics.

Recommended citation: F. Bjelonic, F. Tischhauser, and M. Hutter, “Towards Bridging the Gap: Systematic Sim-to-Real Transfer for Diverse Legged Robots,” arXiv preprint arXiv:2509.06342, 2025.
Download Paper | Download Bibtex