HorizonRobotics/HoloMotion

What it solves

HoloMotion provides an end-to-end framework for humanoid robot control, enabling robots to track complex human motions and follow sparse-to-whole-body commands. It bridges the gap between diverse motion capture data (from VR, inertial, or vision sources) and real-world robot execution, allowing for high-performance teleoperation and motion imitation without requiring per-user training.

How it works

The project uses a reference-conditioned Mixture-of-Experts (MoE) Transformer architecture. It standardizes various input sources into a shared SMPL-based representation through its HoloSMPL module. The HoloRetarget component then converts these motions into specific robot references. The system supports real-time inference, allowing the policy to run directly on-robot or in simulation to control humanoid joints based on either full-body motion sequences or sparse keypoint commands (e.g., pelvis and wrists).

Who it’s for

This project is designed for robotics researchers and engineers working on humanoid locomotion, teleoperation, and whole-body control. It is particularly useful for those looking to deploy pretrained motion-tracking policies on hardware like the G1 humanoid or those needing a standardized pipeline to process motion data for robot training.

Highlights

• Unified Pipeline: Covers the entire workflow from motion data ingestion and retargeting to policy training and real-robot deployment.
• High-Performance Inference: Optimized for real-time control, achieving 300+ FPS on-robot and supporting sequence-level training.
• Flexible Control: Supports multiple command modes, including pelvis-only and multi-point keypoint tracking, within a single policy.
• Cross-Platform Compatibility: Provides tools to integrate various motion capture inputs and supports deployment across simulation environments and physical hardware.