Developers introduced the Light-O1 model for controlling humanoid robots

The development team has released Light-O1 — a specialized text-to-humanoid model designed for generating sequences of robot actions. The system is focused on creating high-precision movements and demonstrates an advantage over existing analogs, such as HY-Motion-1.0 and Kimodo, according to the results of internal benchmarks.

The model's architecture outputs data in the format of 138-dimensional action arrays at a frequency of 20 FPS. Such performance allows for smooth control of humanoid system limbs in real-time. The model has been tested and showed a result of 38 points, taking 13 hours to complete the tasks.

For integration, Light-O1 supports work through a local CLI, Python API, and web console. The software is optimized for deployment on Linux-based operating systems with a CUDA 13 environment. The project's source code is open to the community, allowing robotics developers to implement the model in their own hardware complexes.

The emergence of such specialized solutions simplifies the process of teaching robots complex motor skills through text instructions. The use of 138-dimensional space of actions allows the model to consider more degrees of freedom than in previous generations of control systems, which is critical for the balance and accuracy of humanoid movements.