Enhancing Planning and Decision Making for Robotic Autonomy - John Lathrop

Why It Matters

Hybrid planning that leverages learned priors can overcome data scarcity, enabling safer, faster deployment of next‑generation autonomous robots in real‑world, unpredictable settings.

Key Takeaways

• Data scarcity hampers robot learning in novel autonomous platforms.
• Model-based planning struggles with high-dimensional dynamics and constraints.
• Deep learning offers fast inference but fails out-of-distribution.
• Hybrid approaches combine learned priors to accelerate planning.
• Spectral expansion decision trees enable real-time planning under disturbances.

Summary

John Lathrop’s Everheart lecture tackled the pressing challenge of planning and decision‑making for robotic autonomy when data are scarce. He illustrated the problem with three real‑world projects—a custom tilt‑jet VTOL aircraft, an autonomous racing car, and a simulated spacecraft‑capture mission—each of which lacks the large datasets that power most machine‑learning pipelines. The talk contrasted two dominant paradigms. Model‑based planning relies on explicit dynamics and optimization but quickly becomes intractable in high‑dimensional, disturbance‑rich environments. Deep‑learning approaches, by contrast, deliver millisecond‑scale inference after offline training, yet they falter when confronted with out‑of‑distribution conditions such as unexpected wind, rain, or novel terrain. Lathrop highlighted a hybrid solution: a spectral‑expansion decision‑tree algorithm that embeds learned priors into the planning process. In a live demo, a quadrotor navigated four targets amid complex flow fields and obstacles, solving the problem in real time on a laptop without prior training data. This example underscored how compact representations can bridge the gap between model‑based rigor and deep‑learning speed. The broader implication is clear: future autonomous systems will need to fuse model‑based control with learned representations to remain robust under data scarcity. By advancing hybrid methods, researchers can expand out‑of‑distribution performance, accelerate real‑time planning, and unlock safe deployment of novel robots in unpredictable environments.

Original Description

Enabling robots to plan complex behaviors in real time, rather than relying on predesigned or offline-learned routines, reduces the need for specialized algorithms and large amounts of task-specific training data. At the same time, learning dynamics models through interaction improves planning accuracy, allowing agents to better enforce constraints and predict future behavior. In this talk, John Lathrop presents recent work at the intersection of reinforcement learning, optimization, and robotic autonomy, demonstrated on ground vehicles, aerial systems, and spacecraft. John introduces a new sampling-based planning algorithm with optimality guarantees for continuous, deterministic, differentiable MDPs, encompassing underactuated nonlinear dynamics and nonconvex rewards. He also presents stability guarantees for a coupled dynamics learning and policy optimization framework, validated in the Indy Autonomous Challenge, a high-performance, safety-critical autonomous racing setting.

About the Series:

The Everhart Lecture Series is a forum encouraging interdisciplinary interaction among graduate students and faculty, the sharing of ideas about research developments, as well as a space to discuss controversies. Everhart Lectures allow for the recognition of individual Caltech student's exemplary presentation and research abilities. Each year, three graduate student are selected as Lecturers and receive a $2000 honorarium and recognition at graduation.

This series is cohosted by the Caltech Graduate Office and Graduate Student Council

This lecture was recorded on April 7th, 2026

Produced in association with Caltech Academic Media Technologies.

©2026 California Institute of Technology