Zhangjie Cao

About

Zhangjie Cao is a robotics and machine learning researcher whose work sits at the intersection of imitation learning, reinforcement learning, and computer vision. He is best known for advancing the frontier of learning from imperfect demonstrations — a critical challenge in real-world robotics where ideal expert data is rarely available. His influential work on confidence-aware imitation learning and adversarial confidence transfer enables robots to extract meaningful policies from suboptimal or failure-prone demonstrations, significantly broadening the practical applicability of imitation learning frameworks. Cao has also made notable contributions to spatiotemporal reasoning, particularly in pedestrian intent prediction, which has garnered over 185 citations and underscores his impact on vision-based autonomous systems. His research on transfer reinforcement learning across homotopy classes and feasibility learning for agents with differing dynamics reflects a consistent drive to make robot learning more generalizable and data-efficient. Further work on weakly supervised correspondence learning addresses the challenge of cross-embodiment knowledge transfer without requiring strictly paired datasets. Collectively, Cao's contributions represent a coherent and impactful research agenda aimed at making autonomous systems more robust, adaptable, and deployable in the messy complexity of the real world.

Research Focus

Key Achievements

Top Papers

1. 1
   Spatiotemporal Relationship Reasoning for Pedestrian Intent Prediction

   185 citations · 2020
2. 2
   Confidence-Aware Imitation Learning from Demonstrations with Varying Optimality

   20 citations · 2021
3. 3
   Transfer Reinforcement Learning Across Homotopy Classes

   17 citations · 2021
4. 4
   Learning From Imperfect Demonstrations From Agents With Varying Dynamics

   16 citations · 2021
5. 5
   Spatiotemporal Relationship Reasoning for Pedestrian Intent Prediction

   10 citations · 2020
6. 6
   Learning from Imperfect Demonstrations via Adversarial Confidence Transfer

   5 citations · 2022
7. 7
   Learning Feasibility to Imitate Demonstrators with Different Dynamics

   5 citations · 2021
8. 8
   Weakly Supervised Correspondence Learning

   4 citations · 2022
9. 9
   Learning from Imperfect Demonstrations from Agents with Varying Dynamics

   2 citations · 2021

Key Collaborators

Contact & Links