IMPACT: Intelligent Motion Planning with Acceptable Contact Trajectories via Vision-Language Models
IEEE International Conference on Robotics and Automation (ICRA) · 2026
Karan Owalekar
Robotics · Planning · Learning
Autonomy Engineer at Field AI
I’m an autonomy engineer at Field AI, where I develop spatial representations and planning systems for deployed robots: combining prior maps with live observations, replanning from execution feedback, and language-guided navigation that runs on the robot.
I’m interested in how robots use knowledge, context and experience to decide what to do and to adapt their behavior. Much of what a robot needs to know is implicit: that a glass is fragile, that a corridor was blocked yesterday, that a request refers to a place it has already mapped. I want to build representations and decision mechanisms that make that knowledge usable at planning time while keeping behavior efficient, reliable and easy to inspect.
Previously I worked with Daniel Seita in USC’s SLURM Lab, where I formulated direction-aware contact costs and developed the contact-aware planner for IMPACT (ICRA 2026, co-first author), co-advised by Daniel Seita and Erdem Bıyık. Before USC I worked on hyperspectral imaging and one-shot learning for methane detection at Tata Consultancy Services.
Research interests
Robot Planning and Decision-MakingRepresentations for Robot ReasoningLearning and Adaptation from Experience
Questions I’m interested in
- How can a robot infer what matters for a decision from context, instead of having every relevant distinction specified in advance?
- How can experience, corrections and interaction change a robot’s later behavior without an engineer handling each new situation by hand?
- How can a robot reason about other agents’ intentions and uncertainty when they affect its own actions?
News
Recent publications
One-shot learning for methane leak detection in remote areas using hyperspectral data
IEEE International Geoscience and Remote Sensing Symposium (IGARSS) · 2023