PACE Sim2Real

Systematic sim-to-real transfer via actuator dynamics identification

License

Apache-2.0

Overview

PACE (Precise Adaptation through Continuous Evolution) is a systematic framework from ETH Zurich's Robotic Systems Lab for identifying actuator and joint dynamics to improve sim-to-real transfer of legged robots. Using only standard joint encoders, PACE identifies motor dynamics, friction, damping, and joint compliance — the parameters that most commonly cause the reality gap. The approach is hardware-agnostic and has been validated on multiple quadruped platforms, significantly reducing sim-to-real performance degradation.

PACE (Precise Adaptation through Continuous Evolution) is a framework from ETH Zurich's Robotic Systems Lab that tackles the sim-to-real gap for legged robots. It provides a systematic approach to identifying actuator and joint dynamics parameters using only standard joint encoders, then uses those parameters to create more faithful simulations.

The core method combines data-driven system identification with CMA-ES evolutionary optimization. By recording simple motion trajectories on real hardware, PACE learns physically meaningful parameters like friction, damping, and actuator response characteristics. These parameters are then applied to simulation environments in NVIDIA Isaac Lab, resulting in dramatically better policy transfer.

PACE has been validated on real quadruped platforms including ANYmal, demonstrating measurable improvements in locomotion performance when transferring RL-trained policies from simulation to hardware. The framework is designed to be platform-agnostic, supporting multiple robot configurations and actuator types.

Why it matters

PACE addresses one of the biggest bottlenecks in deploying learned locomotion policies: the reality gap between simulation and physical hardware. By providing a systematic, data-driven method to identify actuator dynamics using only standard joint encoders, it makes high-fidelity sim-to-real transfer accessible without expensive sensor setups.

Problem it solves

Reinforcement learning policies trained in simulation often fail on real robots because simulators poorly model actuator dynamics, friction, and joint behavior. PACE solves this by learning physically meaningful dynamics parameters from real hardware data, then applying them back to simulation for dramatically improved transfer fidelity.

Key takeaways

• Uses CMA-ES evolutionary optimization to identify actuator and joint dynamics parameters
• Requires only standard joint encoders — no additional sensors needed
• Integrates directly with NVIDIA Isaac Lab for seamless sim-to-real workflows
• Supports multiple robot platforms and actuator types out of the box
• Developed and validated by ETH Zurich's Robotic Systems Lab on real quadrupeds

Who it is for

• Legged robotics researchers working on locomotion and sim-to-real transfer
• RL practitioners deploying learned policies from simulation to physical robots
• Robotics engineers needing accurate actuator models for simulation fidelity
• NVIDIA Isaac Lab users looking to improve their sim-to-real pipeline

Limitations

• Requires NVIDIA Isaac Lab (and Isaac Sim 5.0+) for full functionality
• Currently focused on legged robots — applicability to other robot types is untested
• GPU requirements for Isaac Sim may limit accessibility on consumer hardware
• Parameter identification requires collecting real hardware motion data first