RoboBoom
caltech· April 14, 2026· 54:57

Enhancing Planning and Decision Making for Robotic Autonomy - John Lathrop

Caltech PhD student John Lathrop discusses combining planning and learning for data-scarce robotic autonomy, with examples from Caltech’s robotics work and the Indy Autonomous Challenge.

In this talk

  1. 1:20

    Why data scarcity challenges robot learning

    Lathrop focuses on robots that cannot rely on large training datasets, including a custom tilt-jet VTOL aircraft, autonomous race cars, and spacecraft simulators. He argues that model-based planning can struggle with high-dimensional problems, while deep learning can fail when conditions differ from its training data. His research combines the two approaches to improve planning efficiency and performance beyond familiar scenarios.

  2. 13:33

    Spectral expansion guides real-time planning

    For a quadrotor navigating targets, obstacles, and wind, Lathrop presents a method that uses local linear models and controllability information to guide Monte Carlo tree search. The approach runs in real time on a laptop without pre-training and was demonstrated with aerial, ground, and spacecraft systems. Its discretization and search budget involve a trade-off: finer planning can better approximate the continuous problem, but takes longer to converge.

  3. 32:47

    Reinforcement learning for autonomous racing

    In the Indy Autonomous Challenge, Caltech’s team uses modified Dallara AV-24 race cars equipped with lidar, radar, cameras, GPS, and onboard compute. Lathrop works on global planning and multi-car passing, and describes a model-based reinforcement learning method designed to improve performance while learning vehicle dynamics. The method aims to ensure expected policy improvement and better dynamics estimates as track data changes, though the team’s car spun out at Laguna Seca after posting a fast qualifying run.

  4. 45:43

    Future work needs theory, algorithms, and hardware

    Lathrop points to multi-agent driving, contact-rich planning, and data-scarce applications such as space missions, disaster response, and undersea operations. He frames progress around theory that clarifies when learning can succeed, algorithms that handle shifting environments, and hardware validation under real limits such as delays and constrained computation. Autonomous spacecraft repair is one example of a challenging future capability.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.