Humanoid control usually means imitating reference motions in a local root frame, which leaves robots vulnerable to drifting from their intended world position and heading. BeJEPA, a goal-conditioned behavior world model, instead learns a representation that ties body configuration to global placement. It is trained with unsupervised reinforcement learning from human motion priors and robot interaction, using a joint-embedding predictive architecture (JEPA) with two predictors: a Dynamic Predictor that models the next state embedding under an action, and a Successor Predictor that summarizes the expected discounted future embeddings for a goal-conditioned policy.
The two predictors are trained jointly so that the Dynamic Predictor grounds the representation in observed physical transitions while the Successor Predictor organizes it around long-term behavior. Policy improvement then uses one physical prediction and one behavioral summary, avoiding long imagined rollouts. In 20-second motion-tracking benchmarks, BeJEPA with online action refinement outperformed all evaluated baselines in global body-position error, reducing it by 61.7% relative to BFM-Zero with world-pose inputs and 58.7% relative to SONIC.
The authors also deployed BeJEPA on a Unitree G1 humanoid robot. There it achieved joint attainment of posture, world position, and heading goals, recovered from the ground, and returned to its goal after disturbances and goal changes. The paper is available on arXiv under the identifier 2610.23127.