Latent Dynamics · 2026-10-19

BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised RL

Liyun Yan, Haoyang Li, Yang Zhang, Jialing Zhu, Zhanxiang Cao, Zihao Zhou, Keqi Zhu, Yizhi Chen, Huang Chuxuan, Shengcheng Fu, Zhong Jiangwei, Qiao Sun, Chen Zhang, Yue Gao

arXiv:2610.23127 PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

BeJEPA is a goal-conditioned humanoid behavior world model learned with unsupervised reinforcement learning from human motion priors and robot interaction. It uses a joint-embedding predictive architecture (JEPA) with a goal-independent Dynamic Predictor for one-step latent dynamics and a goal-conditioned Successor Predictor summarizing discounted future state embeddings. The composed model guides control without long imagined rollouts, and supports online action refinement.

Why it matters

For world-coordinate humanoid control, tracking posture in a root frame can leave the robot displaced after disturbances. The authors’ Dynamic+Successor decomposition aims to learn a shared representation of body configuration and world placement, so the policy can recover the target global pose (position, heading, and posture) using physical transition prediction plus goal-conditioned long-term behavior prediction.

Method

Learn a shared latent representation encoding body configuration, world position, and heading via a JEPA-style joint-embedding predictive architecture.
Use a goal-independent Dynamic Predictor for one-step latent dynamics (predict z_{t+1} from (z_t,a_t)) and a goal-conditioned Successor Predictor that summarizes discounted future state embeddings under a goal-conditioned policy.
Improve a goal-conditioned off-policy actor by composing one physical prediction with one behavioral summary (no long imagined rollouts), and optionally refine actions online with a gradient-based latent MPC using Dynamic.

Limitation

The authors’ one-step Dynamic predictor “should not be interpreted as a uniformly accurate long-range simulator,” with prediction errors growing over recursive rollout horizons (including drift in root position and heading).

Abstract (from arXiv)

Humanoid control requires coordinating whole-body posture with position and heading in the world frame. Mainstream controllers imitate trajectories in the root frame, but struggle to regain the target world pose after displacement. We introduce BeJEPA, a goal-conditioned Behavior World Model learned through unsupervised reinforcement learning from human motion priors and robot interaction. BeJEPA learns a shared representation of body configuration and world placement through a joint-embedding predictive architecture (JEPA). A Dynamic Predictor predicts the next state embedding under an action, while a Successor Predictor summarizes a goal-conditioned policy's expected discounted future state embeddings in a single vector. During joint training, Dynamic grounds the initial representation in observed physical transitions, while Successor increasingly organizes it around long-term behavior. Their composition guides policy improvement through one physical prediction and one behavioral summary, without long imagined rollouts. In 20-second motion-tracking experiments, BeJEPA with online action refinement using Dynamic outperforms all evaluated baselines in global body-position error, reducing it on AMASS by 61.7% relative to BFM-Zero with world-pose inputs and 58.7% relative to SONIC. Physical deployment on a Unitree G1 demonstrates joint attainment of posture, world position, and heading goals, recovery from the ground, and closed-loop return after disturbances and goal changes. Project page: https://ylyem9x.github.io/BeJEPA-project.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-16 · Indexed by Pub-AI: 2026-10-19 UTC

preprint