Horizon Robotics open-sources HoloMotion-1 for full-body robot control

HoloMotion-1 is an open-source "cerebellum" model designed for humanoid robots, enabling full-body motion control at roughly 300 FPS directly on edge hardware.

While traditional robot controllers usually cap out at a few million parameters and struggle with complex body coordination, HoloMotion-1 scales the policy network to 400 million parameters. Despite its size, it keeps single-step inference down to 3.32 milliseconds. That translates to around 300 FPS, easily beating the standard 50Hz requirement for physical robot control loops.

What it does

The project focuses on what Horizon Robotics calls "Imitate Any Pose"—giving a robot the generic ability to track diverse human motions from real-world data.

To train the model, the team aggregated motions recovered from internet videos, optical motion capture, VR teleoperation, and inertial sensors. All of that feeds into a unified retargeting pipeline to teach the policy how to balance and move like a human.

Tested directly on the Unitree G1 humanoid robot, HoloMotion-1 demonstrated zero-shot transfer without extra fine-tuning. The robot managed dynamic dance routines from web clips, crawled, sat down, performed martial arts high kicks, and followed live teleoperation inputs from a VR controller.

How it works under the hood

Running a 400M-parameter policy in real time requires more than brute force.

HoloMotion-1 uses an MoE-Transformer architecture. When reference poses and robot proprioception arrive, a reference-conditioned router activates only a sparse set of expert networks. Combined with sequence-level optimization and KV-cache mechanisms, this keeps compute overhead remarkably low.

The release includes a complete end-to-end toolchain covering data processing, retargeting, training, evaluation, and physical deployment. Pretrained motion tracking and velocity tracking checkpoints are ready out of the box, with upcoming optimizations targeted at the RDK S100 edge platform offering 80 to 128 TOPS.

Why it matters

For robotics teams and researchers working on bipedal locomotion, HoloMotion-1 offers a massive shortcut past the tedious work of building motion retargeting and baseline policies from scratch.

Historically, robot control relied on narrow, task-specific models that broke whenever you asked for a new posture. HoloMotion-1 shows that scaling up policy capacity and feeding it diverse internet video data works for robot motor skills just like it does for text. It is still an early imitation model that will need deeper perception for rugged terrain, but open-sourcing a practical 300 FPS policy is a huge win for the humanoid developer community.