LeCAR-Lab/BFM-Zero

What it solves

BFM-Zero addresses the challenge of creating versatile, promptable control policies for humanoid robots. Instead of training separate models for every specific task, it provides a foundation model that can be prompted to perform various behaviors—such as motion tracking, goal-reaching, or reward-based tasks—using a unified latent space.

How it works

The project uses unsupervised reinforcement learning to train a behavioral foundation model. By processing large datasets of human motion, the model learns to map complex movements into a compact latent representation (z). During inference, users can provide different prompts or goals, and the model computes the corresponding latent vector to guide the humanoid's movement in a simulator like Isaac Sim or MuJoCo. The resulting policies can be exported as ONNX files for deployment.

Who it’s for

This project is intended for robotics researchers and engineers working on humanoid locomotion, reinforcement learning, and embodied AI. It is particularly useful for those looking to build general-purpose robot controllers that can adapt to different tasks without requiring task-specific retraining.

Highlights

• Promptable Control: Enables a single model to handle diverse tasks like motion tracking and goal-reaching via latent space manipulation.
• Unsupervised Learning: Leverages large-scale motion data to learn robust movement primitives without needing manual reward engineering for every behavior.
• Simulation Support: Fully integrated with Isaac Sim and MuJoCo for training and evaluation.
• Deployment Ready: Includes scripts to export trained policies to ONNX format for cross-platform compatibility.
• Modular Inference: Provides dedicated pipelines for tracking, goal-reaching, and reward-based task execution.