Embodied Navigation
Learning robust navigation from visual observations, language instructions, and complementary expert skills.
PEKING UNIVERSITY · GALBOT · EPIC LAB
Jiawei Chen 陈佳伟
Learning to see, reason, and act.
I am an undergraduate in Computer Science at Peking University, and I am currently at Galbot and EPIC Lab. I will continue my studies as a Ph.D. student at Peking University’s EPIC Lab under the supervision of Prof. He Wang, Prof. Li Yi, and Zhizheng Zhang.
My research focuses on embodied AI, particularly visual navigation and vision-language-action models. I am interested in how robots build useful representations of the world, anticipate what comes next, and turn that understanding into reliable actions.
Research
Representations & World Models
Building spatial and predictive representations that support perception, reasoning, and action.
Reliable Robot Learning
Connecting vision-language-action models with physical reasoning, action evaluation, and recovery.
Selected Publications
* Equal contribution
RepresentNav: See What I Need, Dream What I Want
A Navigation Foundation Model on a Self-Constructed Representation Space
Learns a navigation-oriented visual representation from complementary vision foundation models, then predicts future states through latent reasoning in the same space.
Paper & code forthcoming
VaSeR: From Evaluation to Recovery
Value-Guided Vision-Language-Action Models
Equips a generative VLA with an action-conditioned value learned from demonstrations, enabling candidate selection and recovery through physical backtracking.
Paper & code forthcoming
SPAN-Nav: Generalized Spatial Awareness for Versatile Embodied Navigation
Grounds navigation reasoning in compact 3D spatial priors learned from RGB video, supporting versatile navigation across diverse environments.
LaST-HD: Learning Latent Physical Reasoning from Scalable Human Data for Robot Manipulation
Aligns human and robot demonstrations in a shared latent physical reasoning space to learn manipulation from scalable human-hand data.
MM-Nav: Multi-View VLA Model for Robust Visual Navigation via Multi-Expert Learning
Unifies reaching, squeezing, and obstacle avoidance in a multi-view VLA through capability-balanced learning from multiple navigation experts.
Get in touch
For research discussions and collaboration, feel free to reach out.
2715768122@qq.com