UnifoLM: Predict the Interaction, Not the Whole World
Unitree's UnifoLM is a 6B-parameter world model trained on 2,500 hours of real robot data that predicts only dynamic pixels and optical flow rather than full frames, enabling efficient robot action prediction. This interaction-centric design reduces computational overhead while maintaining high fidelity for embodied control. In this episode of Embodied AI 101, we explore "UnifoLM: Predict the Interaction, Not the Whole World". We break down the research, methodology, and real-world implications for robotics, AI, and physical intelligence. Embodied AI 101 covers the latest research at the…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.