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

cs.ROOct 16, 2026

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

Authors: Liyun Yan, Haoyang Li, Yang Zhang, Jialing Zhu, Zhanxiang Cao, Zihao Zhou, Keqi Zhu, Yizhi Chen,

Organizations: RL2 Lab, Shanghai Jiao Tong University · Shanghai Innovation Institute · Zhejiang University · Tongji University ·

Abstract

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.

Figures & tables

Figure 1: BeJEPA: a goal-conditioned Behavior World Model for humanoid control. Left: our goal is to jointly control world position, heading, and whole-body posture. Center: Dynamic and Successor learn physical consequences and long-term behavior in a shared latent space through unsupervised RL. Right: the Dynamic Predictor refines actions online.

Figure 2: Overview of the BeJEPA architecture. Dynamic predicts the observed next latent, while Successor learns discounted continuation under the collection goal. Both objectives update the shared encoder. The actor learns through their composition, tracing its objective back to the current action. Slashed paths denote stop gradient;

z_{\mathrm{sf}}

denotes a predicted successor feature.

	AMASS				LAFAN
Method	MPJPE ${}_{G}\!\downarrow$	MPJPE ${}_{L}\!\downarrow$	Root err. $\downarrow$	Wrist err. $\downarrow$	MPJPE ${}_{G}\!\downarrow$	MPJPE ${}_{L}\!\downarrow$	Root err. $\downarrow$	Wrist err. $\downarrow$
BFM-Zero (world)	573.4	97.1	0.801	32.4	797.1	89.3	1.020	25.8
SONIC	531.5	49.7	0.571	15.3	887.5	77.8	0.971	23.6
Humanoid-GPT	832.4	115.7	1.107	27.5	1433.1	127.9	1.720	29.7
BeJEPA	278.0	82.3	0.377	25.9	584.5	81.9	0.716	24.3
BeJEPA-MPC (Balanced)	219.5	63.9	0.255	16.3	543.8	75.3	0.639	20.1

Table 1: Whole-body tracking. MPJPE is in mm. Root err. is

e_{xy}+0.01e_{\mathrm{yaw}}

(m, degrees); wrist err. is

e_{\mathrm{pos},L}+0.5e_{\mathrm{rot},L}

(cm, degrees). Bold and underline indicate the best and second-best results.

Figure 3: Tracking trade-offs across execution priorities and temporal alignment on AMASS. (a) Six errors normalized by BeJEPA (1.0); outward is better on a common linear scale. (b) Mean joint-tracking lag over

\pm 500

ms shifts (20 ms steps).

	Root $xy$ (m)		Root yaw ( ∘ )
Model	Min.	Mean	Min.	Mean
BFM-Zero (world)	0.450	0.984	8.28	46.26
Merged predictor	0.666	1.348	6.41	35.81
Split, no Dynamic encoder gradient	0.645	1.231	4.76	34.10
Split, no Successor encoder gradient	0.221	0.885	3.19	25.41
Dynamic-only MPC	0.675	1.004	11.30	82.23

Table 2: Dynamic–Successor ablation on goal inference. Lower errors are better.

Figure 4: World geometry within a shared body-state representation. Eighty body states are crossed with random world poses: $x,y\in[-1,1]$ m and yaw in $[-\pi,\pi]$ . Within-body PCA (PC ρ ) subtracts body means before projection; colors show $x$ , $y$ , and heading.

Figure 5: Motion organization and linear state readouts in the latent space. (a) t-SNE with root $x=y=\psi=0$ ; colors match the illustrated motion families. (b) Linear state readouts at the original world poses.

Figure 6: World-anchored control on the physical G1. (a) Recovery after a kick. (b) Standing up from the ground. (c) Sequential goal switching. Gold rings identify the floor reference marks in (a,c).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	$H$	$R$	$N$	$\epsilon$	$\eta$	$\lambda_{a}$	$\lambda_{\Delta}$
Balanced	16	8	8	0.20	0.080	0.10	0.030
Root $xy$	24	8	8	0.24	0.080	0.03	0.010
Wrist rotation	3	3	6	0.30	0.150	0.01	0.005

Appendix

Table 3: MPC settings for the reported execution priorities.

H

: prediction horizon;

R

: residual horizon;

N

: gradient updates. Bounds

\epsilon

and step sizes

\eta

are in normalized action coordinates. All settings use per-step costs with

\gamma=0.98

Network	Width	Depth	Members	Parameters (M)
State / goal encoder $E_{\phi}$	1,024	6	1	7.12
History Encoder $E_{\eta}^{\mathrm{obs}}$	512	4	1	1.44
Dynamic $D_{\nu}$	2,048	3	2	27.44
Successor $S_{\xi}$	2,048	6	2	53.57
Actor $\pi_{\theta}$	2,048	6	1	31.64
Wrist readout $H_{\zeta}^{w}$	256	4	1	0.27

Appendix

Table 4: Network architecture and size. Depth denotes the main-network hidden-layer count. The CPR/AUX critics use a two-branch design similar to BFM-Zero [ Li et al., 2026 ] ; each input branch adds four residual blocks, explaining their larger parameter count than Successor. Parameter counts include all ensemble members and exclude target-network copies and normalization buffers. M denotes one million parameters.

Parameter	Value	Parameter	Value
Parallel environments	$4{,}096$	Replay capacity	$5.12\times 10^{6}$
Batch size	$4{,}096$	Updates / collection step	$16$
Discount $\gamma$	$0.98$	Target EMA $\tau$	$0.005$
Encoder learning rate	$10^{-5}$	Predictor learning rate	$3\times 10^{-4}$
Actor learning rate	$3\times 10^{-4}$	Observation encoder LR	$3\times 10^{-4}$
Critic learning rate	$10^{-4}$	Discriminator LR	$10^{-5}$

Appendix

Table 5: Core pretraining parameters.

Figure 7: Tracking and fixed-goal errors over training. (a,b) Global and local tracking MPJPE. (c,d) Mean position and heading errors from the separate fixed-goal evaluation. The two split ablations stop the indicated predictor’s gradient to the encoder while retaining both predictors; shading and markers highlight early learning.

Figure 8: Encoder-gradient alignment, weighted norms, and prediction losses. (a) Gradient cosine similarity on the shared encoder. (b) Gradient norms after applying loss coefficients. (c) Unweighted prediction losses averaged over latent coordinates and ensemble members. Faint lines show averages over each logging interval; solid lines show a nine-record moving average. Panels (b,c) use logarithmic vertical axes and share the legend in (c).

Figure 9: Root-position, heading, and joint-angle errors during fixed-goal execution. Lines show mean errors; shading shows the interquartile range. The early approach is expanded horizontally, with markers at $0$ , $0.5$ , $1$ , $1.5$ , $2$ , $5$ , and $10$ s. The illustrated postures receive the same position and heading commands.

Time (s)	1.5	2	5	10
Standing	12.5	87.5	100.0	100.0
Arms out	0.0	87.5	100.0	100.0
Shallow squat	25.0	75.0	100.0	100.0

Appendix

Table 6: Simultaneous goal acquisition (%). Cumulative fraction satisfying root-position error

\leq 0.2

m, yaw error

\leq 15^{\circ}

, and joint RMS

\leq 10^{\circ}

together for

0.5

s. This measures acquisition; the later curves in Figure 9 show holding behavior.

Figure 10: Feature and root-pose prediction errors versus Dynamic rollout horizon. Markers evaluate

1,2,4,8,16,32,

and

64

steps, with

20

ms per step. (a) Feature MSE; the dashed line fits the mean across executions as a linear function of elapsed time. (b,c) Root-position and heading errors, including readout error. Each colored curve averages over starting points within an execution.

Figure 11: Planar root trajectories and base headings for circularly arranged goals. Targets lie on a

1.5

m circle. Open markers and arrows denote target position and yaw; filled markers denote the displayed endpoints. Short, equal-length arrows show base heading along the path, not velocity. We show the arms-out posture with the lowest mean combined alignment error across conditions, tracing each path to its best simultaneous position, heading, and joint alignment within the evaluation window.

Humanoid Robot Control

Humanoid robot control develops methods for coordinating bipedal balance, locomotion, and whole-body interaction with people and objects.

See all 206 papers

Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning
Sep 3, 2026Muyuan Liu, Yue Huang, Zheng Liang +1World Models for Robotics World Model-Based Planning
GENISOM AI, Beijing, China
PhyLatent: Learning Dynamics-Relevant Representations for JEPA World Models
Aug 6, 2026Xi Zeng, Haojie Ren, Ziying Song +2Representation Learning World Model Learning
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore · The University of Sheﬃeld, Sheﬃeld, United Kingdom · School of Artiﬁcial Intelligence (School of Software), Yanshan University, Qinhuangdao, China
Is Forward Prediction Enough? Physical State Grounding for JEPA World Models
Aug 7, 2026Haodong Yan, Jiaguan Zhu, Mingyuan Jia +12Physical Reasoning World Models for Robotics
1The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · 2COCO Matrix

Sep 3, 2026J/K move · Enter open · S save

Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning

Muyuan Liu, Yue Huang, Zheng Liang, Xiang Gao

GENISOM AI, Beijing, China

World Models for Robotics World Model-Based Planning Joint-Embedding Predictive Architecture Action-Conditioned World Models Robot Task Planning Latent-Space Planning

Open paper AskPDF arXiv

Action-conditioned JEPA world models enable planning toward visually specified goals without reconstructing future pixels, yet latent prediction alone does not explicitly encourage the learned representations to retain information relevant to robotic control. We introduce an end-to-end JEPA world model that augments latent prediction with inverse dynamics (IDM) and state alignment (SA). While inverse dynamics discourages latent collapse and makes latent transitions informative of the actions that produced them, state alignment grounds consecutive representations in their associated physical configuration and motion. Across four benchmark tasks, our model attains the highest success rates on TwoRoom (100%), PushT (98%), and OGBench-Cube (87%), while performing comparably to LeWorldModel on Reacher. Our ablation further shows that adding state alignment consistently improves planning success over IDM alone across all four tasks. Although LeWorldModel, our primary baseline, attains higher average straightening on OGBench-Cube, transition-subspace analysis shows that its transition energy is concentrated in a substantially lower-dimensional subspace. Our state-aligned model exhibits a higher effective transition dimension than LeWorldModel and improves planning over IDM alone, supporting state alignment as an effective complement to inverse dynamics for robotic planning.

Abstract

Humanoid Robot Control

Explore similar work

Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning

PhyLatent: Learning Dynamics-Relevant Representations for JEPA World Models

Is Forward Prediction Enough? Physical State Grounding for JEPA World Models