Jonathan DeCastro

I am a Staff Research Scientist and research lead at the Toyota Research Institute in Cambridge, MA, where I lead the Shared Decision-Making project: model-based reinforcement learning that assists humans in time-critical, high-speed shared-control settings. My research gives learning-based systems logical, game-theoretic, and control-theoretic structure so they can behave safely and trustworthily around people, especially when human data is scarce. I received my Ph.D. in 2017 from Cornell University as an affiliate of the Verifiable Robotics Research Group, where I worked with Prof. Hadas Kress-Gazit on automated, correct-by-construction controller synthesis for complex robots.

My research interests lie in the intersection of reinforcement learning, game theory, control theory and formal methods. I have a deep interest in the following topics:

• Human-AI collaboration and shared autonomy in time-critical, high-speed settings.
• Game-theoretic and sequential decision-making models for human-robot interaction.
• Correct-by-construction synthesis of continuous controllers from temporal logic specifications.
• Learning from scarce human data: preference learning, cognitive-bias inference, and personalization.
• Algorithms for falsification of interesting system properties, such as safety.
• Application of formal synthesis, verification, and falsification to robotics applications of all kinds.

Learning an Interpretable Logic Monitor for Risk-Aware and Socially-Compliant Trajectory Planning
Xiao Li, Jonathan DeCastro, Cristian-Ioan Vasile, Sertac Karaman and Daniela Rus
International Journal of Robotics Research, 2025.

Personalizing Driver Safety Interfaces via Driver Cognitive Factors Inference
Emily Sumner, Jonathan DeCastro, et al.
Nature Scientific Reports, 14(1):18058, 2024.

A Safe Preference Learning Approach for Personalization With Applications to Autonomous Vehicles
Ruya Karagulle, Nikos Aréchiga, Jonathan DeCastro and Necmiye Ozay
IEEE Robotics and Automation Letters, 9(5):4226–4233, 2024.

Abstracting Road Traffic via Topological Braids: Applications to Traffic Flow Analysis and Distributed Control
Christoforos Mavrogiannis, Jonathan DeCastro and Siddhartha Srinivasa International Journal of Robotics Research, 43(9):1299–1321, 2024.
[pdf | bibtex | link]

CARPAL: Confidence-Aware Intent Recognition for Parallel Autonomy
Xin Huang, Stephen McGill, Jonathan DeCastro, Luke Fletcher, John Leonard, Brian Williams and Guy Rosman
Robotics and Automation Letters, 6(3):4433–4440, 2021.
[arXiv]

Vehicle Trajectory Prediction using Generative Adversarial Network With Temporal Logic Syntax Tree Features
Xiao Li, Guy Rosman, Igor Gilitschenski, Cristian-Ioan Vasile, Jonathan A. DeCastro, Sertac Karaman and Daniela Rus
Robotics and Automation Letters, 6(2):3459–3466, 2021.
[link]

DiversityGAN: Diversity-Aware Vehicle Motion Prediction via Latent Semantic Sampling
Xin Huang, Stephen McGill, Jonathan DeCastro, Brian Williams, Luke Fletcher, John Leonard and Guy Rosman
Robotics and Automation Letters, 5(4):5089–5096, 2020.
[arXiv]

Reactive Mission and Motion Planning while Avoiding Dynamic Obstacles
Javier Alonso-Mora, Jonathan DeCastro, Vasumathi Raman, Daniela Rus and Hadas Kress-Gazit
Autonomous Robots, 42(4):801–824, April, 2018.
[pdf | bibtex | link]

Automated Generation of Dynamics-Based Runtime Certificates for High-Level Control
Jonathan DeCastro, Ruediger Ehlers, Matthias Rungger, Ayca Balkan and Hadas Kress-Gazit
Journal of Discrete Event Dynamical Systems: Special Topical Issue on Formal Methods in Control, 27(2):371-405, 2017.
[pdf | bibtex | link]

Synthesis of Nonlinear Continuous Controllers for Verifiably-Correct High-Level, Reactive Behaviors
Jonathan DeCastro and Hadas Kress-Gazit
International Journal of Robotics Research, 34(3):378-394, 2015.
[pdf | bibtex | link]

Learning to Plan, Planning to Learn: Adaptive Hierarchical RL-MPC for Sample-Efficient Decision Making
Toshiaki Hori, Jonathan DeCastro, Deepak Gopinath, Arun Balachandran and Guy Rosman
L4DC 2026: Learning for Dynamics & Control Conference.

Timing the Message: Language-Based Notifications for Time-Critical Assistive Settings
Y. C. Hsu, Jonathan DeCastro, Alexander Silva and Guy Rosman
AAMAS 2026: International Conference on Autonomous Agents and Multiagent Systems.

Computational Teaching for Driving via Multi-Task Imitation Learning
Deepak Gopinath, Xiongyi Cui, Jonathan DeCastro, Emily Sumner, et al.
ICRA 2025: IEEE International Conference on Robotics and Automation, pp. 7019–7027.

Safety with Agency: Human-Centered Safety Filter with Application to AI-Assisted Motorsports
Donggun D. Oh, Justin Lidard, Haimin Hu, et al., Jonathan DeCastro, Guy Rosman, Naomi Leonard and Jaime F. Fisac
RSS 2025: Robotics: Science and Systems.

Enhancing Trust Repair in Driving: The Role of AI Emotional Expression and Perceived Control
Xiaozhen Hu, Lucas Dees, Deepak Gopinath, Alexander Silva, Jonathan DeCastro, Guy Rosman and Emily Sumner
ITSC 2025: IEEE 28th International Conference on Intelligent Transportation Systems.

Think Deep and Fast: Learning Neural Nonlinear Opinion Dynamics from Inverse Dynamic Games for Split-Second Interactions
Haimin Hu, Jonathan DeCastro, Deepak Gopinath, Guy Rosman, Naomi E. Leonard and Jaime Fisac
ICRA 2025: IEEE International Conference on Robotics and Automation.

Estimating Cognitive Biases with Attention-Aware Inverse Planning
Sreejan Banerjee, Daphne Cornelisse, Deepak Gopinath, Emily Sumner, Jonathan DeCastro, Guy Rosman, et al.
NeurIPS 2024: Advances in Neural Information Processing Systems, 38:14524–14550.

Dreaming to Assist: Learning to Align with Human Objectives for Shared Control in High-Speed Racing
Jonathan DeCastro, Alexander Silva, Deepak Gopinath, Emily Sumner, Tucker Balch, Lucas Dees and Guy Rosman
CoRL 2024: 8th Conference on Robot Learning.

Incorporating Logic in Online Preference Learning for Safe Personalization of Autonomous Vehicles
Ruya Karagulle, Necmiye Ozay, Nikos Aréchiga, Jonathan DeCastro and Andrew Best
HSCC 2024: 27th ACM International Conference on Hybrid Systems: Computation and Control, pp. 1–11.

Blending Data-Driven Priors in Dynamic Games
Justin Lidard, Haimin Hu, Asher Hancock, Ziyi Zhang, Albert Gimó Contreras, Vikash Modi, Jonathan DeCastro, Deepak Gopinath, Guy Rosman, Naomi Leonard, Mario Santos and Jaime Fernández Fisac
RSS 2024: Robotics: Science and Systems.

Classification of Driving Behaviors Using STL Formulas: A Comparative Study
Ruya Karagulle, Nikos Aréchiga, Jonathan DeCastro and Necmiye Ozay
FORMATS 2022: 20th International Conference on Formal Modeling and Analysis of Timed Systems. Warsaw, Poland.
[link]

Outlier-Robust Inverse Reinforcement Learning and Reward-based Detection of Anomalous Driving Behaviors
Dan Li, Mohamad Louai Shehab, Zexiang Liu, Nikos Aréchiga, Jonathan DeCastro and Necmiye Ozay
ITSC 2022: 25th International Conference on Intelligent Transportation Systems. Macau, China.
[link]

Implicit Multi-Agent Coordination at Unsignalized Intersections via Topological Inference
Christoforos Mavrogiannis, Jonathan A. DeCastro and Siddhartha S. Srinivasa
WAFR 2022: Fifteenth Workshop on the Algorithmic Foundations of Robotics. College Park, MD, USA.
[arXiv]

Analyzing Multiagent Interactions in Traffic Scenes via Topological Braids
Christoforos Mavrogiannis, Jonathan A. DeCastro and Siddhartha S. Srinivasa
ICRA 2022: International Conference on Robotics and Automation. Philadelphia, PA, USA.
[arXiv]

Learning A Risk-Aware Trajectory Planner From Demonstrations Using Logic Monitor
Brandon Araki, Xiao Li, Kiran Vodrahalli, Jonathan Decastro, Micah Fry and Daniela Rus
CoRL 2021: 5th Annual Conference on Robot Learning. London, UK.
[link]

The Logical Options Framework
Brandon Araki, Xiao Li, Kiran Vodrahalli, Jonathan Decastro, Micah Fry and Daniela Rus
ICML 2021: 38th International Conference on Machine Learning. Virtual Conference.
[link]

Differentiable Logic Layer for Rule Guided Trajectory Prediction
Xiao Li, Guy Rosman, Igor Gilitschenski, Jonathan DeCastro, Cristian-Ioan Vasile, Sertac Karaman and Daniela Rus
CoRL 2020: 4th Annual Conference on Robot Learning. Virtual Conference.
[link | video]

Behaviorally Diverse Traffic Simulation via Reinforcement Learning
Shinya Shiroshita, Shirou Maruyama, Daisuke Nishiyama, Mario Ynocente Castro, Karim Hamzaoui, Guy Rosman, Jonathan DeCastro, Kuan-Hui Lee and Adrien Gaidon
IROS 2020: IEEE/RSJ International Conference on Intelligent Robots and Systems. Virtual Conference.
[arXiv]

Discovering Avoidable Planner Failures of Autonomous Vehicles using Counterfactual Analysis in Behaviorally Diverse Simulation
Daisuke Nishiyama, Mario Ynocente Castro, Shirou Maruyama, Shinya Shiroshita, Karim Hamzaoui, Yi Ouyang, Guy Rosman, Jonathan DeCastro, Kuan-Hui Lee and Adrien Gaidon
ITSC 2020: The 23rd IEEE International Conference on Intelligent Transportation Systems. Virtual Conference.
[arXiv]

Interpretable Policies from Formally-Specified Temporal Properties
Jonathan DeCastro, Karen Leung, Nikos Aréchiga and Marco Pavone
ITSC 2020: The 23rd IEEE International Conference on Intelligent Transportation Systems. Virtual Conference.
[pdf | bibtex]

Counterexample-Guided Safety Contracts for Autonomous Driving
Jonathan DeCastro, Lucas Liebenwein, Cristian-Ioan Vasile, Russ Tedrake, Sertac Karaman and Daniela Rus
WAFR 2018: The 13th International Workshop on the Algorithmic Foundations of Robotics. Mérida, Mexico.
[pdf | bibtex]

Compositional and Contract-based Verification for Autonomous Driving on Road Networks
Lucas Liebenwein, Wilko Schwarting, Cristian-Ioan Vasile, Jonathan DeCastro, Javier Alonso-Mora, Sertac Karaman and Daniela Rus
ISRR 2017: International Symposium on Robotics Research. Puerto Varas, Chile.
[pdf | bibtex | video]

Nonlinear Controller Synthesis and Automatic Workspace Partitioning for Reactive High-Level Behaviors
Jonathan DeCastro and Hadas Kress-Gazit
HSCC 2016: 19th ACM International Conference on Hybrid Systems: Computation and Control. Vienna, Austria.
[pdf | bibtex | video]

Collision-Free Reactive Mission and Motion Planning for Multi-Robot Systems
Jonathan DeCastro, Javier Alonso-Mora, Vasumathi Raman, Daniela Rus and Hadas Kress-Gazit
ISRR 2015: International Symposium on Robotics Research. Sestri Levante, Italy.
[pdf | bibtex | video]

Dynamics-Driven Adaptive Abstraction for Reactive High-Level Mission and Motion Planning
Jonathan DeCastro, Vasumathi Raman and Hadas Kress-Gazit
ICRA 2015: IEEE International Conference on Robotics and Automation. Seattle, WA, USA.
[pdf | bibtex]

Guaranteeing Reactive High-Level Behaviors for Robots with Complex Dynamics
Jonathan DeCastro and Hadas Kress-Gazit
IROS 2013: IEEE/RSJ International Conference on Intelligent Robots and Systems. Tokyo, Japan.
[pdf | bibtex]

HMIway-env: A Framework for Simulating Behaviors and Preferences to Support Human-AI Teaming in Driving
Deepak Gopinath, Jonathan DeCastro, Guy Rosman, Emily Sumner, Allison Morgan, Shabnam Hakimi and Simon Stent
HCIS 2022: Human-centered Intelligent Services Safety and Trustworthiness at CVPR 2022. New Orleans, LA, USA.
[link]

Certified Control for Self-Driving Cars
Daniel Jackson, Jonathan DeCastro, Soonho Kong, Dimitrios Koutentakis, Angela Leong Feng Ping, Armando Solar-Lezama, Mike Wang and Xin Zhang
DARS 2019: 4th Workshop on the Design and Analysis of Robust Systems. New York, NY, USA.
[pdf]

Better AI through Logical Scaffolding
Nikos Arechiga, Jonathan DeCastro, Soonho Kong and Karen Leung
FoMLAS 2019: 2nd Workshop on Formal Methods for ML-Enabled Autonomous Systems. New York, NY, USA.
[arXiv]

On the Strengths and Weaknesses of Data for Open-set Embodied Assistance
Pradyumna Tambwekar, Alexander Silva, Deepak Gopinath, Jonathan DeCastro, Xiongyi Cui and Guy Rosman
arXiv 2026
[arXiv]

Spatio-Temporal Retrieval-based Priors for Adaptive Computational Teaching in Driving
Deepak Gopinath, Xiongyi Cui, Jonathan DeCastro, Arun Balachandran and Guy Rosman
arXiv 2026
[arXiv]

Boundary Sampling to Learn Predictive Safety Filters via Pontryagin’s Maximum Principle
Jason Dallas, Thomas Lew, Jenny Talbot, Jonathan DeCastro, Somil Bansal and John Subosits
arXiv 2026
[arXiv]

Specification-Guided Data Aggregation for Semantically Aware Imitation Learning
Ameesh Shah, Jonathan DeCastro, John Gideon, Guy Rosman and Sanjit Seshia
In review, 2023

NashFormer: Leveraging Local Nash Equilibria for Semantically Diverse Trajectory Prediction
Justin Lidard, Oswin So, Yanxia Zhang, Jonathan DeCastro, Xiongyi Cui, Xin Huang and Guy Rosman
arXiv 2023
[arXiv]

Learning Latent Traits for Simulated Cooperative Driving Tasks
Jonathan A. DeCastro, Deepak Gopinath, Guy Rosman, Emily Sumner, Shabnam Hakimi and Simon Stent
CoRR 2022
[arXiv]

Dynamics-Based Reactive Synthesis and Automated Revisions for High-Level Robot Control
Jonathan DeCastro, Ruediger Ehlers, Matthias Rungger, Ayca Balkan, Paulo Tabuada and Hadas Kress-Gazit
CoRR 2016
[pdf | arXiv]

Shared Autonomy & Human-AI Teaming

I lead TRI's Shared Decision-Making project, developing model-based reinforcement learning (Dreamer) to assist humans in time-critical, high-speed shared-control settings. We deployed a shared-control system on a real-time, human-in-the-loop simulator and demonstrated it live with human drivers, including on-track sessions at Thunderhill Raceway, extending the work to a VR-based driving rig and TRI's Global Research Innovation Platform (GRIP) drivable test bench.

Publications: [Dreaming to Assist (CoRL 2024)] [Safety with Agency (RSS 2025)] [Timing the Message (AAMAS 2026)]

Structure for Learning: Formal Methods Meet Reinforcement Learning

A throughline across my work is giving learned policies verifiable, interpretable structure through logic monitors, signal temporal logic, control-barrier functions, and safety contracts. This keeps learning-based systems safe and trustworthy under distribution shift and when human interaction data is scarce.

Publications: [Learning an Interpretable Logic Monitor (IJRR 2025)] [Boundary Sampling for Predictive Safety Filters (2026)] [Counterexample-Guided Safety Contracts (WAFR 2018)]

Game-Theoretic & Sequential Decision-Making

Dynamic games, Nash equilibria, and opinion dynamics for split-second human-robot interaction, together with hierarchical RL-MPC for sample-efficient planning under partial observability of a human partner's intent. These models let an autonomous agent reason about how its actions influence a human partner's decisions, and vice versa, in real time.

Publications: [Think Deep and Fast (ICRA 2025)] [Blending Data-Driven Priors in Dynamic Games (RSS 2024)] [Learning to Plan, Planning to Learn (L4DC 2026)]

Learning from Scarce Human Data

Preference learning, cognitive-bias inference, and computational teaching methods that personalize to individual humans without requiring large interaction datasets. The goal is to adapt quickly to a person's preferences and capabilities from a handful of interactions rather than extensive prior data collection.

Publications: [Personalizing Driver Safety Interfaces (Nature Sci. Reports 2024)] [Estimating Cognitive Biases with Attention-Aware Inverse Planning (NeurIPS 2024)] [Computational Teaching for Driving (ICRA 2025)]

Formally-Verified Contracts for Safe Deployment of Self-Driving Cars

Verifying the safety of autonomous vehicles is paramount, but formal verification of autonomous driving decision systems is generally computationally intractable. With collaborators at MIT, I developed a framework for constructing safety contracts, sufficient conditions on a controller that guarantee safety with respect to probabilistic traffic models, and demonstrated the approach scaling across a range of multi-car scenarios and road topologies.

Publications: [Liebenwein, et. al. (ISRR 2017)] [DeCastro & Liebenwein, et. al. (WAFR 2019)] [video]

Automatic Synthesis of High-Level, Reactive Controllers for Nonlinear Systems

This work developed an approach for automatically synthesizing libraries of low-level controllers that, together, guarantee a high-level, temporal-logic specification for robots with complex dynamics. It included transferring the specification into sound mathematical structures for the low-level controllers, along with abstraction refinement and specification repair to keep controllers realizable under environment uncertainty.

Low-Complexity Synthesis for Multi-Robot Scenarios with Dynamic Obstacles

Controller synthesis for multi-robot systems can become intractable as the number of robots and dynamic obstacles grows. I developed a high-level synthesis scheme using a local planner with collision-avoidance guarantees that coordinates multiple robots without observing the global behavior of dynamic obstacles, preserving correctness while avoiding this exponential blowup.

Publications: [DeCastro, et. al. (ISRR 2015)] [video]