Papers
3
Total Citations
22
H-Index
3
About
DiJia Su’s research lies at the intersection of autonomous navigation, behavior prediction, and sample-efficient reinforcement learning (RL), with a strong emphasis on deploying robust models in real-world, safety-critical settings. In his most cited work, Su tackles the coordinate-frame gap in autonomous driving by proposing a distillation method for scene-centric motion forecasting, enabling efficient and accurate predictions of agent behavior—a foundational contribution to safe motion planning. He also introduced MUSBO, a model-based batch optimization framework that incorporates uncertainty regularization to achieve sample-efficient RL under deployment constraints, directly addressing the impracticality of continuous online learning in fields like healthcare and robotics. Earlier in his career, Su developed a hierarchical controller for autonomous robot navigation that combines high-level Q-learning with low-level visual servoing to avoid randomly moving obstacles. With over 20 citations across his top papers, Su’s work demonstrates a clear trajectory from foundational robotics to cutting-edge, deployment-aware AI systems. His research is particularly notable for bridging the gap between theoretical RL advances and practical, constrained real-world applications.
Research Focus
Key Achievements
Top Papers
- 1
- 2
- 3