Papers
5
Total Citations
154
H-Index
4
About
Kelvin Xu is a robotics and machine learning researcher whose work sits at the intersection of reinforcement learning, imitation learning, and autonomous robotic manipulation. His research tackles some of the most persistent challenges in deploying RL agents in real-world settings, particularly the design of reward functions and the reduction of human intervention during training. Xu's most influential contribution, "Unsupervised Perceptual Rewards for Imitation Learning" (2017, 124 citations), introduced a framework for automatically generating reward signals from visual observations, significantly reducing the need for hand-engineered reward functions — a longstanding bottleneck in practical RL deployment. This work has become a foundational reference in the imitation learning community. His subsequent research has pushed toward increasingly autonomous robotic systems. Notable projects include reset-free RL methods that allow dexterous manipulation agents to learn continuously without human intervention, and substep-guided approaches for mastering complex, contact-rich tasks with multi-fingered robotic hands. His 2021 work on formalizing autonomous RL further established theoretical grounding for this emerging paradigm. Across his publications, Xu has demonstrated a consistent commitment to bridging the gap between laboratory RL research and practical, real-world robotic applications.
Research Focus
Key Achievements
Top Papers
- 1Unsupervised Perceptual Rewards for Imitation Learning124 citations · 2017
- 2
- 3
- 4Unsupervised Perceptual Rewards for Imitation Learning5 citations · 2016
- 5Autonomous Reinforcement Learning: Formalism and Benchmarking3 citations · 2021