Event
BeJEPA: Goal-Conditioned Humanoid Behavior World Model Paper Released
1 reports · 1 sources
Liyun Yan and other researchers proposed and released the BeJEPA paper, a goal-conditioned humanoid robot behavior world model learned via unsupervised reinforcement learning from human motion priors and robot interactions, using a Joint Embedding Predictive Architecture (JEPA) to learn a shared representation of body configuration and world placement. The paper reports that in 20-second motion tracking experiments, BeJEPA combined with Dynamic-based online motion refinement outperformed all evaluated baselines on global body position error, reducing it by 61.7% relative to BFM-Zero (with world pose input) and by 58.7% relative to SONIC on AMASS. In physical deployment on the Unitree G1, the model achieved joint attainment of posture, world position, and heading goals, rising from the ground, and closed-loop recovery after perturbations and goal changes.
Report timeline
- BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised Reinforcement Learning
The paper proposes BeJEPA, a goal-conditioned humanoid behavior world model learned through unsupervised reinforcement learning from human motion priors and robot interaction, using a joint-embedding predictive architecture (JEPA) to learn a shared representation of body configuration and world placement.
arXiv cs.RO ·