CAREER: Learning for Robot Motion Planning using Physics Priors

Humans possess an intuitive understanding of physics that enables them to coordinate complex movements and interactions with their environments in real time, often with little or no training or guidance. In contrast, robots are not yet capable of this and generally need substantial precomputation, pretraining, expert demonstrations, or learning through trial and error. This project aims to enhance robots’ ability to interpret and use physics for planning their movements and interactions in constrained and unknown environments. The goal is to enable robots to learn rapidly with minimal expert supervision and to plan their movements in real time. Such capabilities are important for robots that must operate in new and unknown real-world settings. For example, home-assistance robots may need to operate in new homes and fetch, carry, clean, or help with daily-life tasks. Similarly, factory robots may need to handle tools and workpieces without extensive reprogramming, and search-and-rescue robots may need to move through damaged or unknown areas. The project will validate the methods on robot navigation and manipulation tasks involving known and unknown environments, tool handling, and home-assistance scenarios. The outcomes of this project will be shared through publications, open-source code, benchmarks, student research opportunities, and educational outreach activities designed to expand access to robotics. This research develops a new class of learning-based methods for robot motion planning using physics priors. In this project, physics priors refer to partial differential equations, or PDEs, that describe robot motion, constraints, and cost-to-go functions. The research is organized around two thrusts. The first thrust identifies suitable PDEs and studies how to solve them to generate robot motion under complex collision and kinematic constraints. The project will design learning-based methods to solve these PDEs with low training cost and fast run-time speed. It also studies how to solve these equations under multimodal manifold constraints, which are important for robot manipulation and locomotion. The second thrust investigates how robots can learn and adapt motion-planning policies online in unknown environments. This is studied in the context of unknown obstacle geometry, unknown dynamics, and novel kinematic constraints. The work focuses on formulating and solving PDEs in real time from exploratory observations and limited interaction experience. It also studies methods that scale to large operational spaces without forgetting previously learned information. The methods are demonstrated on robot navigation and manipulation tasks in known and unknown environments, including tool handling and home-assistance tasks. These research outcomes are integrated into educational activities for undergraduate and graduate students. The project also develops K-12 educational activities that use real-time tools developed from the project outcomes. These activities help students explore robot motion and manipulation planning and understand how these ideas apply to robots that operate and assist in real-world settings. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.