A major breakthrough in robot motion intelligence is reshaping how humanoid robots move and respond in real time. Chinese firm Horizon Robotics has released HoloMotion-1, a 4-billion-parameter AI model designed to control whole-body humanoid robot movements with unprecedented speed and fluidity. The system can perform real-time inference at 300 frames per second on edge devices, meaning robots can now make movement decisions almost instantly without relying on cloud computing.

What Makes This Different From Previous Robot Control Systems?

Until now, humanoid robots relied on much smaller AI models to control their movements. Previous systems used motion control models with only millions or tens of millions of parameters, which limited how smoothly and naturally robots could move. HoloMotion-1 scales that up dramatically to 4 billion parameters, giving robots access to vastly more learned movement patterns.

The key innovation lies in how the system learns. Instead of relying solely on motion capture data recorded in controlled studio environments, HoloMotion-1 trains on a diverse mix of movement data, including curated motion capture recordings, proprietary motion data created by Horizon Robotics, and movements reconstructed from real-world videos. This broader training approach helps robots handle movements they have never directly encountered before and adapt when sensors don't work perfectly.

The system uses a Transformer-based neural network, a type of deep learning architecture that excels at understanding sequences over time. This is a significant upgrade from older, simpler neural network designs that struggled with long and complex motion patterns. To keep the system fast enough for real-time use, Horizon Robotics employs a Mixture-of-Experts approach, which activates only the most relevant parts of the model at each moment, saving computing power.

How Are Robots Using This Technology Right Now?

• Direct Real-World Transfer: Horizon Robotics tested HoloMotion-1 directly on a Unitree G1 humanoid robot without any additional training on real-world data, and the robot successfully performed movements it had never been trained on in hardware, including dancing, crawling, sitting, and martial arts-style kicks.
• Live Human Control: The system supports real-time teleoperation through motion capture suits and virtual reality controllers, allowing humans to pilot robots that closely follow their movements with stable, responsive tracking.
• Edge Device Deployment: The robot's onboard computer handles all movement calculations at 200 to 300 decision cycles per second, while the robot's physical motion system runs at 50 cycles per second to maintain smooth, stable movement.

The results demonstrate what researchers call a "sim-to-real" transfer, meaning the robot learned movement patterns in simulation and successfully applied them to physical hardware without additional adjustment. This is a major milestone because it reduces the time and cost needed to deploy new robot behaviors.

What Does This Mean for the Future of Humanoid Robots?

This development arrives at a critical moment in robotics. Industry experts have predicted that bipedal robots will begin assisting tradespeople within two years, and that robots could eventually gain more advanced "brains" capable of independent reasoning. HoloMotion-1 represents a major step toward that vision by solving one of the hardest problems in robotics: making robots move naturally and responsively in unpredictable real-world environments.

The broader context matters too. Horizon Robotics has positioned HoloMotion-1 as the first step in a four-stage roadmap for humanoid robot control: Imitate Any Pose, Follow Any Command, Move on Any Terrain, and Control Any Robot Type. By completing the first stage, the system lays groundwork for future improvements that could enable robots to handle increasingly complex tasks.

Meanwhile, other companies are racing to build competing systems. Unitree, another major robotics firm, has launched UniStore, described as the world's first app store for humanoid robots, allowing owners of the G1 humanoid to download skill packages like dance routines, martial arts moves, and walking styles. This shift toward developer ecosystems suggests the industry is moving beyond raw hardware competition toward software and AI capabilities.

"HoloMotion-1 represents a significant advance in scalable humanoid robot control and edge AI deployment," Horizon Robotics stated in describing the system's capabilities.

Horizon Robotics, company statement

The timing is significant because AI researchers are increasingly optimistic about near-term breakthroughs. Jack Clark, co-founder of Anthropic, recently predicted that bipedal robots will help tradespeople within two years, and that by the end of 2028, AI systems will be able to design their own successors. While such predictions carry uncertainty, they reflect genuine momentum in the field.

For robotics companies and researchers, the challenge now is scaling these advances beyond controlled demonstrations into real-world deployments. HoloMotion-1 shows that the technical barriers are falling, but questions remain about cost, reliability, and the practical economics of deploying humanoid robots at scale. The next phase will test whether this progress translates into robots that can reliably perform useful work in factories, warehouses, and other demanding environments.