leggedrobotics/pace-sim2real

PACE: A systematic approach for sim-to-real transfer of legged robots, identifying actuator and joint dynamics with standard joint encoders.

What it solves

PACE addresses the "sim-to-real gap"—the discrepancy between how a robot behaves in a physics simulator and how it behaves in the real world. This gap often prevents controllers trained in simulation from working effectively on physical hardware, particularly for legged robots where actuator and joint dynamics are complex.

How it works

The framework uses a data-driven approach to align simulation with reality. It collects measured data from the physical robot and uses the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to optimize and fit physically meaningful dynamics parameters. By estimating actuator and joint dynamics directly from this data, it creates a more accurate simulation model that can be used for robust locomotion training.

Who it’s for

It is designed for robotics researchers and engineers working with legged robotic systems who use NVIDIA Isaac Lab for simulation and need to transfer learned behaviors to real hardware.

Highlights

  • Data-Driven Identification: Estimates actuator and joint dynamics directly from measured real-world data.
  • Evolutionary Optimization: Employs CMA-ES for precise parameter fitting.
  • Isaac Lab Integration: Built to integrate seamlessly with NVIDIA Isaac Lab's task and environment conventions.
  • Platform Agnostic: Supports multiple robot platforms and various actuator types.

Related

  • Project
    isaac-sim/IsaacLab

    A GPU-accelerated open-source framework built on NVIDIA Isaac Sim for simplifying robotics research in reinforcement learning, imitation learning, and motion planning.

  • Project
    fan-ziqi/rl_sar

    A framework for simulation verification and physical deployment of reinforcement learning policies for quadruped, wheeled, and humanoid robots.

  • Dispatch
    Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model – OpenAI 2016

    OpenAI researchers presented a method that transfers simulation-learned control policies to real robots by learning a deep inverse dynamics model to map simulated next states to appropriate real-world actions, reducing the sim-to-real gap without requiring accurate dynamics models.

  • Project
    ACEsuit/mace

    MACE is an open‑source PyTorch library for equivariant graph‑neural‑network interatomic potentials. It provides fast training/evaluation, multi‑GPU and Apple‑Silicon support, optional CUDA acceleration, and a catalog of pre‑trained foundation models for materials and organic chemistry. Install via `pip install mace-torch`, train with the `mace_run_train` CLI (or YAML configs), and evaluate with `mace_eval_configs`. Documentation, Colab tutorials, and Weights & Biases integration are included. The project is stable on the v0.3 line while a major v1.0 rewrite is in progress.

  • Project
    nvidia-isaac/WBC-AGILE

    A reinforcement learning framework based on NVIDIA Isaac Lab for training humanoid loco-manipulation policies with validated sim-to-real transfer capabilities.