JEPLO: Joint-Embedding Predictive Learning for
LiDAR-Based Legged Locomotion

Qihao Yuan* Yixuan Qiu* Ziyu Cao† Ming Cao* Kailai Li†
*University of Groningen †Linköping University
Code arXiv YouTube Hardware Setup Dataset

Abstract

We present JEPLO (Joint-Embedding Predictive learning for legged LOcomotion), a single-stage learning framework for mapping-free, LiDAR-based perceptive locomotion for legged robots. We introduce a proprio-exteroceptive JEPA (PE-JEPA) world model to learn predictive egocentric terrain representations from onboard observations, including raw LiDAR scans. A concurrent JEPA-teacher-student (CJTS) pipeline is further proposed to train a locomotion policy informed by JEPA latent representations in simulation using deep reinforcement learning with a simple reward formulation. The framework achieves successful sim-to-real transfer, enabling omnidirectional traversal of diverse terrains, including long staircases and high boxes, with lightweight onboard computation. Evaluations demonstrate greater robustness than existing perceptive locomotion frameworks, particularly under degraded perception caused by occlusion, sparsity and noise. Further analysis validates JEPLO's ability to retain task-relevant information under these challenging conditions.

System and Training Pipeline

Omnidirectional Locomotion

Mapping-free perceptive locomotion. Meshes are only used for visualization.

Forward upstairs and downstairs

Sideway upstairs and downstairs

Backward upstairs and downstairs

In-place rotation

Repeated Trials

Mapping-free perceptive locomotion. Meshes are only used for visualization.

High box (53 cm, repeated 5 times)

Mixed terrain (repeated 5 times)

Locomotion in Darkness

Mapping-free perceptive locomotion. Meshes are only used for visualization.

Forward upstairs and backward downstairs

Box jumping

Perception Degradation

Mapping-free perceptive locomotion. Meshes are only used for visualization.

Cross Occlusion

Sparsity (single scan)

Box Climbing

Mapping-free perceptive locomotion. Meshes are only used for visualization.

High box (53 cm)

Consecutive jumps

Angled jumps

Outdoor Tests

Long stairs

Outdoor Step

Garage

Narrow space (2x speedup)

Oooops

Citation

@article{yuan2026jeplo,
  title={JEPLO: Joint-Embedding Predictive Learning for LiDAR-Based Legged Locomotion}, 
  author={Yuan, Qihao and Qiu, Yixuan and Cao, Ziyu and Cao, Ming and Li, Kailai},
  journal={arXiv preprint arXiv:2609.15770},
  year={2026}
}