AI

HoloMotion: Horizon’s Open Whole-Body Humanoid Controller Hits 300 FPS

Robb Harlan 4 min read

Horizon Robotics has open-sourced HoloMotion, a foundation model for whole-body humanoid motion control. Interesting Engineering’s write-up on HoloMotion-1 describes a large MoE Transformer “cerebellum” aimed at real-time imitation and teleoperation on edge hardware, with GitHub releases through v1.4 (July 16, 2026).

What it does

HoloMotion is built for motion tracking and whole-body imitation rather than full language-to-task VLA stacks. The project page and README describe a reference-conditioned MoE Transformer trained on large motion corpora — the v1.3 notes scale parameters from tens of millions toward 0.4B and motion data from tens of hours toward 2,000+ hours, with policy inference moving from roughly 100 toward ~300 FPS.

Interesting Engineering reports on-device inference around 200–300 control cycles per second on a Unitree G1’s onboard computer, with the low-level motion loop at about 50 Hz. The same coverage says the policy transferred zero-shot from simulation onto G1 hardware without extra real-world fine-tuning, showing dance, crawl, sit, and kick-style motions plus live VR / MoCap teleoperation.

Unitree G1 performing dance, crawl, kick, and teleop box carry with HoloMotion
Zero-shot and teleop motion demos on Unitree G1. Source: Interesting Engineering / Horizon Robotics.

Stack and roadmap

GitHub documents three user paths: offline motion replay, live teleoperation, and full train-your-own. HoloSMPL unifies VR, inertial, optical, and vision capture; HoloRetarget (v1.4) is claimed at 3,000+ FPS on RTX 4090 for data generation and 300+ FPS fully on-robot for teleop retargeting.

The roadmap is explicit:

  • v1.x Any Pose — motion imitation (done)
  • v2.x Any Command — language / task-conditioned motion (next)
  • v3.x Any Terrain — planned
  • v4.x Any Embodiment — planned

Technical report: arXiv:2605.15336. Code, Docker deployment docs, and Hugging Face model collection are public.

HoloMotion MoE policy architecture and latency charts
MoE policy architecture and training/inference latency charts from the open repo. Source: HorizonRobotics/HoloMotion on GitHub.

A Human’s Take

Whole-body tracking that actually runs at control rates is the boring layer factories need before fancy task models matter. I’m less interested in the martial-arts kick reels than in whether HoloRetarget and the Docker path let a lab stick this on a G1 without a six-month systems project. Command following and terrain are still roadmap — judge the project on those when they ship, not on the dance clips.

Sources