Research
Our work is organized around three questions that span the autonomy stack: how autonomous systems acquire competence through learning and interaction, how they coordinate with one another, and how we ensure safety and reliability when acting in the physical world. A fourth thrust cuts across these, grounding our methods in high-stakes applications with societal-scale impact.
Representative Papers
Thrust 01
Robot Learning
The emergence of internet-scale, broadly capable Foundation Models (FMs) presents an opportunity to fundamentally rethink how autonomous systems are designed, deployed, and operated. Trained on vast and diverse datasets, these models encode broad priors about the world and have driven breakthroughs in vision, language, and multimodal reasoning. Yet deploying AI in physical embodiments introduces challenges that differ fundamentally from those encountered in purely digital domains.
We develop the tools and algorithms that enable Physical AI by grounding general-purpose robot policies in the dynamics and constraints of the physical world. This includes learning representations for the rich multimodal streams robots perceive; advancing vision-language-action models, world models, and reasoning models for robust decision-making; and integrating learned components with model-based methods that provide structure, safety, and efficiency. These systems must also operate within the constraints of onboard compute while learning effectively from the limited embodiment-specific data available in the real world.
Thrust 02
Coordinated Autonomy
Autonomous systems are becoming integral to modern society, increasingly embedded in critical infrastructure, transportation, manufacturing, scientific discovery, and human-centered services. As these systems proliferate, they will rarely operate in isolation. Fleets of autonomous vehicles will coordinate to move people through cities, constellations of spacecraft will jointly monitor the Earth, and teams of robots will collaborate with one another and alongside people. In these settings, the object of design is not the individual robot, but the collective system.
We develop learning, optimization, and control methods for decision-making across interacting agents. Our research addresses the fundamental challenges that arise in large-scale multi-agent systems, including limited communication, partial observability, real-time computational constraints, and environments in which agents must cooperate, compete, or do both simultaneously.
Our goal is to advance the foundations of autonomous decision-making at the intersection of AI, control theory, and large-scale optimization. We are particularly interested in coordination algorithms that explicitly model interactions with the broader societal systems in which autonomous agents operate, such as transportation networks and the electric grid, enabling the joint optimization of mobility, energy, and other shared resources. Ultimately, we seek to develop methods that empower operators and policymakers to balance competing objectives, including performance, sustainability, resilience, and equity, while ensuring the safe, efficient, and reliable operation of autonomous systems at scale.
Representative Papers
Representative Papers
- •Vision Foundation Model Embedding-based Semantic Anomaly Detection · IEEE ICRA, 2025
- •Do Vision-Language-Action Models Mean What They Say? On the Role of Faithfulness in Embodied Reasoning · arXiv, 2026
- •Towards Robust Spacecraft Trajectory Optimization via Transformers · IEEE Aerospace Conference, 2025
Thrust 03
Physical AI Safety
Delivering reliable, general-purpose autonomy presents fundamental challenges for AI systems. Autonomous systems must operate safely in complex, open-world environments, often under conditions that cannot be anticipated during design or training. Unlike purely digital applications, failures in the physical world carry tangible consequences, including harm to people, damage to critical infrastructure, and mission failure. Yet today’s learning-based systems, despite their remarkable capabilities, remain susceptible to brittle generalization under distribution shift, making their behavior difficult to predict, verify, and trust in safety-critical settings.
We develop methods that combine the adaptability of modern AI with the reliability of model-based autonomy. Rather than replacing classical planning and control, our goal is to integrate learning-based components in ways that preserve the safety and robustness required for real-world deployment. This includes runtime monitors that detect semantic and distributional anomalies before they propagate into unsafe behavior; guardrails that constrain learned policies to provably safe or verifiable action sets; and architectures that leverage the broad knowledge encoded in foundation models to interpret unfamiliar situations, reason about their implications, and synthesize safe recovery strategies instead of defaulting to conservative fail-stop behavior.
A central theme of our work is trustworthy embodied reasoning. As foundation models increasingly become the decision-making engines of autonomous systems, it is essential that their reasoning be grounded in the robot’s observations, dynamics, and objectives, rather than relying on plausible but unfaithful explanations. We develop methods for interpretable and faithful reasoning that expose the evidence behind a system’s decisions, quantify uncertainty, and enable humans to verify why an autonomous system selected a particular course of action. By combining grounded reasoning with principled safety mechanisms, we aim to build autonomous systems that are not only more capable, but also predictable, transparent, and worthy of deployment in mission-critical environments.
Thrust 04
Societally-impactful Autonomy
Representative Papers
- •Semantic Trajectory Generation for Goal-Oriented Spacecraft Rendezvous · AIAA SciTech, 2026
- •Space-LLaVA: a Vision-Language Model Adapted to Extraterrestrial Applications · IEEE Aerospace Conference, 2025
- •Reproducibility in the Control of Autonomous Mobility-on-Demand Systems · IEEE T-RO, 2025
Ultimately, our goal is to expand the frontier of autonomous systems by tackling problems in domains where reliability is non-negotiable, human supervision is limited, and the consequences of failure are significant. These environments push existing methods to their limits, exposing fundamental challenges in perception, decision-making, coordination, and verification that must be solved before autonomous systems can be broadly trusted.
We pursue this research through applications that both inspire and validate our work. In aerospace autonomy, we develop algorithms for spacecraft, aerial systems, and other mission-critical platforms operating in uncertain and communication-constrained environments. In future mobility, we study autonomy across scales, from decision-making for individual autonomous vehicles to the coordination of fleets and their interaction with transportation and energy infrastructure, to improve the safety, efficiency, and sustainability of next-generation mobility systems. More broadly, we develop the foundations of general-purpose autonomy for robots and autonomous systems that must perceive, reason, and act safely in complex, dynamic, and open-world environments.
By grounding our research in these challenging application domains, we aim to advance the fundamental science of autonomy while enabling technologies that deliver meaningful societal impact.
A full list of our work, including preprints and theses, is on the Publications page.