HorizonRobotics/HoloMotion

What it solves

HoloMotion addresses the challenge of scaling humanoid whole-body control. It enables robots to robustly imitate diverse human motions and track references in real-time, while remaining flexible enough to work across different hardware configurations without needing a separate policy for every single assembly.

How it works

The system uses a reference-conditioned Mixture-of-Experts (MoE) Transformer trained on large-scale motion data. It includes a standardized representation called HoloSMPL that unifies diverse motion capture sources (VR, inertial, optical, and vision) for a retargeting process (HoloRetarget). This pipeline allows the system to bridge the gap between raw motion data, policy learning, simulation, and real-world deployment.

Who it’s for

• Robot researchers looking for a foundation for humanoid motion control.
• Developers wanting to perform offline motion replay or live teleoperation of humanoid robots.
• ML engineers aiming to train custom motion policies from their own datasets.

Highlights

• High-performance inference: Supports real-time teleoperation at 300+ FPS on-robot and data generation at 3,000+ FPS on GPU.
• Hardware versatility: A single policy can run across 66 unique compatible configurations of the Unitree G1 robot (varying heads, bodies, and hands).
• Scalable architecture: Uses an MoE Transformer that has scaled from 60M to 0.4B parameters.
• Unified data pipeline: Standardizes multiple motion capture inputs into a single shared representation.