Unitree Robotics has published UnifoLM-WLA-1.0, a new version of its general-purpose embodied foundation model, releasing training code, inference code and model weights openly. The company reports state-of-the-art results across several benchmarks among open-source models.

A foundation model in robotics plays a role similar to a foundation model in language. Rather than programming a robot separately for each task, the model is trained broadly enough to be adapted to many of them, converting what the robot sees and is told into the movements it makes.

What distinguishes this release is scope rather than scale. At six billion parameters it is not a large model by current standards. But a single model handles both desktop manipulation, meaning work done with the hands from a fixed position, and whole-body mobile manipulation, meaning the robot moves through a space and then manipulates something once it arrives. Those have historically been separate control problems handled by separate systems.

The model also generalizes across end effectors. A policy trained with one type of gripper or hand can transfer to another without being retrained from the ground up, which matters because hardware varies widely between robots and between tasks.

This follows UnifoLM-VLA-0, released in January, which unified vision, language and action in a single model. Unitree has also developed UnifoLM-X2-1.0, a real-time world model used in the autonomous humanoid demonstrations it showed in September.

Open weights change who can build on this. When training code and weights are published rather than licensed, the capability stops being tied to one manufacturer's hardware and becomes a starting point available to any team working with humanoid platforms.