Unitree Robotics humanoid and foundation model development

Unitree Robotics has introduced the UnifoLM-X2-1.0, which the company claims is the first fully autonomous humanoid robot operating on a real-time world model. This system utilizes a world-action foundation model to predict physical interactions and force dynamics, allowing for movement planning without remote controls or pre-programmed sequences.

Initial demonstrations featured two humanoid robots engaging in nearly two minutes of autonomous combat, performing punches, kicks, and defensive maneuvers. The company described this as “combat-level autonomy.” Subsequent demonstrations focused on the Unitree G1 humanoid robot, which utilized the UnifoLM-X2-1.0 architecture to face a human fighter in a boxing ring. The robot operated independently, using a predictive approach to recognize its environment and anticipate an opponent's movements. During these tests, the G1 demonstrated the ability to adjust foot positioning to dodge strikes and respond with combinations of attacks.

At the IFA technology fair in Berlin, Unitree showcased the G1 participating in a boxing match against a human trainer. The demonstration emphasized the robot’s ability to evaluate evolving combat scenarios to proactively adjust posture and maintain balance. However, observers noted that while the G1 shows high agility, technical challenges regarding latency and computational load persist.

Expanding its software ecosystem, Unitree has announced the open-sourcing of UnifoLM-WLA-1.0, a new generation of general-purpose humanoid robot foundation models. This 6B parameter model was trained on 2,500 hours of real-machine data and is capable of performing 64 different tasks across various robot bodies. The model utilizes embodied reasoning, future change prediction, and action learning to enhance spatial perception and decision-making.