Ioannis Kantaros
Washington University in St Louis, MO·Engineering·Early-career
Teams of ground and aerial robots could one day handle deliveries, search-and-rescue missions, disaster relief and factory work in places they have never seen before. To make decisions under that uncertainty, robots increasingly rely on reinforcement learning (RL), a type of AI that learns control strategies from trial and error.
- Fields
Engineering
Computer Science
- Topics
robotics
multi-robot systems
multi-agent reinforcement learning
- Grants
- 3 known grants, through Apr 2029
What they work on
- Train multi-agent reinforcement learning (MARL) control policies for robot teams and mathematically certify their robustness against adversarial, high-dimensional sensor inputs, testing them on aerial and ground robot testbeds and photorealistic simulators.
- Assess and improve how well MARL policies hold up when messages between robots are corrupted or when robots unexpectedly fail or are removed from the team.
- Determine whether robust equilibria, stable outcomes no agent gains by leaving alone, exist in non-cooperative multi-robot settings with self-interested agents, and develop ways to compute them.
- Develop multi-robot planners that, given a high-level task, jointly design cooperative manipulation plans for reshaping blocked or disconnected environments and motion plans for completing the task.
- Integrate these interactive planners with active perception, so robots sense in ways that reduce uncertainty about object shape and identity and can interact with surroundings safely.
- Speed up safe deep reinforcement learning for mobile robots whose goals are written in temporal logic, a formal language for tasks that unfold over time, such as visiting places in order while avoiding hazards.
- Develop methods to safely transfer learned robot controllers to new temporal logic missions and environments, and to verify neural-network controllers compositionally, checking them piece by piece.
More about the research
RL controllers are powerful but hard to trust. They can fail when sensors are noisy, when messages between robots are corrupted, or when the surroundings block the task. This research asks how to build robot autonomy that comes with mathematical safety guarantees and that can adapt to the physical world. It combines formal methods (mathematical tools for proving a system meets its specification), machine learning and control theory.
Why it matters
Autonomous robot teams are attractive for dangerous or labor-intensive work such as search-and-rescue, disaster response, manufacturing, infrastructure inspection and environmental monitoring. They are hard to deploy where mistakes are costly because their AI-based controllers lack safety guarantees. Algorithms that can be certified as robust, and that let robots work around obstacles in unfamiliar places, could make these systems more dependable for emergency responders, manufacturers and delivery services.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
CAREER: Interactive Planning for Perception-enabled Robot Teams in Unknown Semantic Environments
NSF · May 2024 – Apr 2029
CPS: MEDIUM: Certified Robust Learning for Multi-Agent Planning and Control
NSF · Jun 2024 – May 2027
CPS: SMALL: Formal Methods for Safe, Efficient, and Transferable Learning-enabled Autonomy
NSF · Apr 2023 – Mar 2027
Closest labs
Navid Azizan Ruhi
Massachusetts Institute of Technology, MA·Engineering·Early-career
Develop Hard-Constrained Neural Networks (HardNet), which use a differentiable projection layer to guarantee input-dependent constraints, and embed them in reinforcement learning to produce certifiably safe policies.
Thinh Doan
University of Texas at Austin, TX·Engineering·Early-career
Develop the theory of two-time-scale stochastic approximation for non-monotone systems, using extrapolation from optimization and singular perturbation from control to keep these noisy iterative methods stable, and prove finite-time complexity bounds.
Vijay Subramanian
University of Michigan, MI·Engineering
Develop robust reinforcement learning algorithms that use model-class knowledge, future predictions and off-policy methods to cope with mismatches and limited data from simulators or offline records.
The school
Washington University in St Louis: 97 labs on PI Pages, $1.2B a year on research, 365 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.