CE-Nav: Flow-Guided Reinforcement Refinement for
Cross-Embodiment Local Navigation

Kai Yang^* Tianlin Zhang^* Zhengbo Wang^* Zedong Chu^†

Xiaolong Wu Yang Cai Mu Xu

AMAP, Alibaba Group

^*Equal contribution ^†Corresponding author

CE-Nav is a learning-based, generalizable local navigation framework for robots. It features a novel two-stage (Imitation Learning-then-Reinforcement Learning) methodology that systematically decouples universal geometric reasoning from embodiment-specific dynamic adaptation, enabling efficient policy transfer across diverse robot morphologies including quadrupeds, bipeds, and quadrotors.

Overview

Overview of the CE-Nav two-stage framework. Stage 1 (Left): A multi-modal, embodiment-agnostic General Expert is trained offline via imitation learning on expert data. Stage 2 (Right): The frozen expert is used as a guiding prior to train a Dynamics-Aware Refiner via online reinforcement learning, allowing it to adapt to a specific robot's dynamics.

Cross-Embodiment Navigation

Go2

MagicDog

Spot

H1

Hummingbird

VelFlow: Multi-Modal Velocity Planning

CE-Nav Real-World Deployment

CE-Nav Indoor

CE-Nav Outdoor

CE-Nav can navigate complex indoor and outdoor environments, reliably avoiding both static and dynamic obstacles while maintaining speeds of up to 1.5m/s to efficiently arrive at its destination.

Citation

If you use CE-Nav in your research, please cite our paper:

@article{yang2025cenav,
  title={{CE-Nav: Flow-Guided Reinforcement Refinement for Cross-Embodiment Local Navigation}},
  author={Yang, Kai and Zhang, Tianlin and Wang, Zhengbo and Chu, Zedong and Wu, Xiaolong and Cai, Yang and Xu, Mu},
  journal={arXiv preprint arXiv:2509.23203},
  year={2025},
  eprint={2509.23203},
  archivePrefix={arXiv},
  primaryClass={cs.RO}
}