nvidia-isaac/WBC-AGILE

A Comprehensive Workflow for Humanoid Loco-Manipulation Learning

What it solves

AGILE simplifies the process of training humanoid robots to perform complex movements (loco-manipulation), such as walking, standing up, and dancing, while ensuring that behaviors learned in a simulator can be successfully transferred to physical robots (sim-to-real).

How it works

Built on NVIDIA Isaac Lab, AGILE uses reinforcement learning to train whole-body control policies. It employs a teacher-student distillation process where a "teacher" policy is trained using privileged information available only in simulation, which is then distilled into a "student" policy that can operate using only the sensors available on the real robot.

Who it’s for

It is designed for robotics researchers and practitioners working with humanoid robots, specifically those using platforms like the Booster T1 and Unitree G1.

Highlights

  • Multi-Robot Support: Validated for sim-to-real transfer on both Booster T1 and Unitree G1 robots.
  • Teacher-Student Distillation: Enables the creation of deployable policies by distilling privileged simulation data into student policies.
  • Cross-Simulator Validation: Includes a framework to transfer policies from Isaac Lab to MuJoCo for further validation.
  • Remote Training: Supports OSMO workflows for large-scale training and sweeps on compute clusters.
  • Integrated Evaluation: Provides motion metrics, HTML reports, and Weights & Biases integration for tracking performance.

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
    zitongbai/legged_lab

    A reinforcement learning extension for Isaac Lab that enables humanoid robots, such as the Unitree G1, to learn complex motions via DeepMimic and Adversarial Motion Priors.

  • Project
    fan-ziqi/rl_sar

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

  • Project
    NVlabs/ProtoMotions

    A GPU-accelerated simulation and learning framework for training physically simulated digital humans and humanoid robots, supporting large-scale motion learning and sim-to-real deployment.

  • Project
    Kaixhin/imitation-learning

    A unified framework for benchmarking and implementing deep imitation learning algorithms like GAIL and RED using SAC as a base, tested on Gym MuJoCo environments.