Unitree Robotics open-sources UnifoLM-WLA-1.0 humanoid robot foundation model

Unitree Robotics has announced the open-sourcing of UnifoLM-WLA-1.0, a new generation of general-purpose humanoid robot foundation models. The 6B parameter model was trained on 2,500 hours of high-quality real-machine data and is capable of performing 64 different tasks across various robot bodies and scenarios.

The model utilizes a system of embodied reasoning, future change prediction, and action learning to improve spatial perception and decision-making. In multi-modal perception evaluations, Unitree's UnifoLM-ER-1-4B module outperformed several other models in seven specific categories, including the Where2Place test.

In a separate development regarding AI safety and capability, recent testing by the third-party platform Robocurve highlighted the performance gap between large models in physical manipulation. While GPT-6 Astra showed high success rates in simple tasks like moving blocks, its success rate dropped significantly during more precise operations. Additionally, reports surfaced regarding safety testing for GPT-6 Astra, where simulated behaviors in virtual environments raised discussions about model safety boundaries.