About Me
Hello and welcome to my website! I am a postdoctoral researcher hosted by Prof. Joel Burdick and Prof. Yisong Yue at the California Institute of Technology (Caltech). I received the B.S. in Mechanical Engineering from Stanford University, and the M.S. and PhD. in Aeronautics from Caltech, advised by Prof. Soon-Jo Chung.
Research Summary
I am broadly interested in planning, machine learning, and dynamical systems with applications in robotics, space autonomy, and self-driving cars. During my Ph.D. and postdoctoral research, I focused on developing algorithms that combine search (predictive reasoning) and learning (pattern recognition) intelligence modalities in robotics. We validated our algorithms on systems that included spacecraft robots, aerial robots, tracked vehicles, self-driving cars, and manipulators. I work on theory, algorithm development, and hardware experiments and use techniques from controls, dynamical systems, machine learning, and optimization. More generally, I am interested in the study and design of sophisticated models of physical intelligence as a dynamical system, and validating these models with impactful robotic experiments.
News
- I am honored to be selected for the William F. Ballhaus Doctoral Prize! (Departmental Thesis Award) (1 or 2 per year at Caltech).
- I am honored to be selected as a 2024 Future Leader in Robotics and AI by Microsoft Robotics and University of Maryland!.
- I am honored to be selected as a RSS Pioneer 2023!.
- Our paper was selected as Best Graduate Student Paper Award in the field of Guidance, Navigation, and Control at the 2023 AIAA SciTech Forum!!
- I enjoyed my summer as a research intern at Motional on the Machine Learning Planning team (2021).
- I am honored to organize the GALCIT Colloquium Series during 2020-2021.
- Our paper was selected as Honorable Mention for Best Paper of IEEE Robotics and Automation Letters 2020 (<1% of all papers)!
Selected Publications
(* denotes equal contribution)
Monte Carlo Tree Search for Dynamical Systems with Spectral Expansion
Benjamin Riviere*, John Lathrop* and Soon-Jo Chung.
Spectral Expansion Tree Search (SETS) is a real-time planning algorithm that uses dynamical systems theory to construct a low-complexity and provably-correct representation of high-dimensional and continuous decision-making problems. Published in Science Robotics December 2024 Paper.
Online Tree-based Planning for Active Spacecraft Fault Estimation and Collision Avoidance
James Ragan, Benjamin Riviere and Soon-Jo Chung.
Safe Fault Estimation with Active Sensing Tree Search (sFEAST) selects actions to excite observations most useful in diagnosing a fault affecting a robot's sensors and actuators, while probabilistically ensuring safety over the planning horizon. Published in Science Robotics August 2024 Paper.
Bayesian Active Sensing for Fault Estimation with Belief Space Tree Search
James Ragan*, Benjamin Riviere* and Soon-Jo Chung.
Partially Observable Monte-Carlo Planning with Marginalized Filtering (POMCPMF) restores strategic exploration in information gathering tree search by computing exact Bayesian transitions, enabled by belief distribution factorization in fault estimation problems. Accepted at AIAA SciTech Forum 2023. Paper. Best Graduate Student Paper Award in the field of Guidance, Navigation, and Control at the 2023 AIAA SciTech Forum.
Neural Tree Expansion for Multi-Robot Planning in Non-Cooperative Environments
Benjamin Riviere, Wolfgang Hoenig, Matthew Anderson and Soon-Jo Chung.
Neural Tree Expansion (NTE) biases tree search with decentralized deep neural network generative models for real-time game-theoretic planning in high dimensional continuous spaces. Accepted at IEEE Robotics and Automation Letters (Robotics and Automation Letters) and International Conference on Intelligent Robots and Systems (IROS) 2021. Arxiv, Code, Video
Global-to-Local Safe Autonomy Synthesis for Multi-Robot Motion Planning with End-to-End Learning
Benjamin Riviere, Wolfgang Hoenig, Yisong Yue and Soon-Jo Chung.
GLAS learns decentralized coordinated policies from centralized data and uses nonlinear stability to enable end-to-end training with safety guarantees. Accepted at IEEE Robotics and Automation Letters and IROS 2020. Arxiv, Code, Video Honorable Mention for Best Paper of IEEE Robotics and Automation Letters 2020
Hybrid Temporal Difference Learning for Adaptive Urban Taxi Dispatch
Benjamin Riviere and Soon-Jo Chung.
H-TD^2 coordinates thousands of urban taxis with a combination of: decentralized Kalman filtering for information sharing, local online reinforcement learning for value estimation, and game-theoretic task assignment for coordination. Accepted at IEEE Transactions on Intelligent Transportation Systems 2021. Arxiv