Mechanical and Civil Engineering Seminar

Mechanical and Civil Engineering Seminar Series

Title: 'Probabilistic robotics' revisited: Risk-sensitive robot planning and control using data-driven geometry and dynamics models

Abstract: Emerging tools such as large language models (LLMs), foundation models, and neural fields promise significant advancements in robot perception and high-level reasoning. However, these benefits have not yet materialized as robust robot behavior in hardware. One reason for this gap is that robotics pipelines (whether model-based or data-driven) are often brittle, failing unpredictably when key modeling assumptions are violated, or when operating far from their training distribution. My work aims to address this brittleness by developing theoretical and algorithmic tools for enabling autonomous systems to reason about risk.

In this talk, I will discuss three of my recent projects that take steps toward robust, risk-sensitive decision-making for robot navigation, manipulation, and control. First, I will present our work on using Neural Radiance Fields (NeRFs) for visual navigation, which enables robots to plan risk-sensitive paths through a cluttered environment using only an onboard RGB camera. Next, I will discuss our recent work on robust grasp optimization for dexterous hands, which uses a gradient-based, bilevel optimization to plan grasps that are robust to disturbances at speeds up to 100x faster than existing methods. Finally, I will present our work on data-driven risk sensitive control which combines stochastic control barrier functions, which provide rigorous, probabilistic safety guarantees, with deep generative disturbance models. We demonstrate our method in both simulation and hardware, where our controller can perform aggressive, safe quadrotor flight with a completely unmodeled and uninstrumented slung load. The goal of my talk is to highlight how risk sensitivity and uncertainty representation provide interesting ways forward for combining rigorous low-level control with modern, learning-based planners and perception.

Bio: Preston Culbertson is currently a postdoctoral scholar at Caltech, collaborating with Prof. Aaron Ames on risk-sensitive robot planning and control. He completed his PhD at Stanford University under Prof. Mac Schwager, focusing on collaborative manipulation and assembly with robot teams. His work seeks to integrate rigorous control methods with cutting-edge, learning-based approaches in robotic planning and perception. Specifically, he is interested in enabling robust robot behavior in hardware for difficult tasks like manipulation, locomotion, and navigation through new methods for uncertainty representation and risk-sensitive robot control and planning. He received the NASA Space Technology Research Fellowship (NSTRF) and was awarded the 'Best Manipulation Paper' at ICRA 2018.

NOTE: At this time, in-person Mechanical and Civil Engineering Lectures are open to all Caltech students/staff/faculty/visitors.