To truly transform our lives, autonomous systems must operate in complex environments shared with other agents. For instance, delivery robots navigate spaces with humans, while warehouse robots must coordinate on shared factory floors. These settings require systematic methods that enable efficient and reliable interactions among multiple agents. In this talk, I will focus on imitation learning in interactive multi-agent domains, with an emphasis on game-theoretic settings. In particular, I will discuss both behavior cloning and inverse reinforcement learning in the context of strategic interactions among agents. Unlike single-agent settings, where behavior can be learned from independent demonstrations, multi-agent environments require reasoning about interdependent decisions, where each agent’s behavior is shaped by others. I will highlight key challenges that arise in this setting, including the coupling between agents’ actions and the difficulty of inferring consistent objectives and policies from demonstrations of interactions. I will then discuss approaches that leverage game-theoretic structure to make imitation learning more tractable and effective in these domains.

Efficient learning in network Markov Decision Processes (MDPs), a fundamental model in multi-agent reinforcement learning, is extremely challenging due to exponential growth of the global state-action space with the number of agents. In this work, utilizing the exponential decay property of network dynamics, we first derive scalable spectral local representations for multiagent reinforcement learning in network MDPs, which induces a network linear subspace for the local Q-function of each agent. Building on these local spectral representations, we design a scalable algorithmic framework for multiagent reinforcement learning in continuous state-action network MDPs, and provide end-to-end guarantees for the convergence of our algorithm. Empirically, we validate the effectiveness of our scalable representation-based approach on two benchmark problems, and demonstrate the advantages of our approach over generic function approximation approaches to representing the local Q-functions.

As autonomous systems increasingly operate alongside humans and other machines, their movements are continuously observed and interpreted by others. An agent’s observed behavior can indirectly reveal hidden information such as intent, capabilities, or knowledge, creating a tight coupling between control actions and the signals conveyed to observers. This talk investigates game-theoretic approaches for designing behaviors that strategically trade off task efficiency against information disclosure in order to influence an observer’s decision-making. In particular, we will discuss an initial set of works focused on adversarial settings, where deceptive and counter-deceptive behaviors emerge naturally from strategic interaction. The talk also addresses computational challenges associated with incomplete-information games, along with algorithmic approaches that enable tractable analysis in physically embodied systems. Overall, this work aims to establish a principled foundation for signaling-aware motion control in interactive multi-agent systems operating under uncertainty.

As autonomous systems scale to decentralized multi-agent settings, agents must make decisions in the presence of others with limited and asymmetric information, across both cooperative and non-cooperative interactions. This raises a fundamental question: how can we model, compute, and learn strategic decisions in such environments? This talk approaches this question from a differentiable game-theoretic perspective. By approximating game-theoretic optimality conditions as differentiable equations, we enable efficient computation and learning of equilibria under different information structures in complex dynamic games. I will present scalable algorithms for solving nonlinear feedback dynamic games with convergence and safety guarantees, along with inverse game-theoretic methods for inferring agents’ objectives and beliefs about others from partial observations. I will then discuss how these components can be combined in a closed loop, enabling agents to align with others or exploit information asymmetry when beneficial. We demonstrate these methods in applications such as advanced air mobility, multi-robot furniture moving, and drone racing, including hardware experiments, where decentralized agents coordinate in real time without direct communication. Overall, these results suggest that differentiable game-theoretic structure enables efficient computation and learning of multi-agent strategies in complex, interactive environments.

Robots operating in crowded human-populated environments must be capable of inferring the interaction constraints, such as collision avoidance specifications, that often underlie multi-agent behaviors. To empower robots to learn interaction constraints, this talk presents an inverse dynamic game-based framework for inferring parametric constraints from multi-agent interaction demonstrations at local Nash equilibria. To recover constraints consistent with the local Nash stationarity of the given demonstrations, we encode the corresponding Karush–Kuhn–Tucker (KKT) conditions within a mixed-integer linear program (MILP). We establish theoretical guarantees that our method learns inner approximations of the true safe and unsafe sets. We also use the recovered interaction constraint information to design motion plans that robustly satisfy the true, a priori unknown constraints despite limited demonstration data. Across simulations and hardware experiments, our method accurately infers constraints from interaction demonstrations and leverages the inferred constraint information to design safe interactive motion plans. We conclude by outlining ongoing and future work on extending the proposed framework to enable active, parameterization-free, real-time, and online interaction constraint inference. The research contributions presented in this talk are the result of collaborations with Zhouyu Zhang, Zheng Qiu, and Dr. Glen Chou.

Robot life isn’t what it used to be. Gone are the days of certainty and comfort brought by industrial cages. From autonomous vehicles to general-purpose humanoids, robots need to work closely with people, learning to gauge our intent, anticipate our actions, and adapt their own decisions swiftly and reliably. The value proposition of modern robotics hinges on seamless interaction, but how can autonomy pipelines tap into human theory of mind and social behavior? This talk argues that dynamic game theory provides a crucial mathematical backbone for human-centered robot decision-making: information structure. We will explore information-based strategies across the spectrum from fully cooperative to fully adversarial games: from pedagogy, where players’ actions can efficiently encode pivotal information through implicit cues, to deception, where agents must hedge against ambiguous or misleading opponent behavior. Examples will span autonomous driving in dense urban traffic, human–robot collaboration, and AI assistants. The talk concludes with an outlook on human-centered intelligence, outlining opportunities and open challenges for the coming years.

Learning enabled services are revolutionizing several engineering domains such as robotics, mobility, energy, and online marketplaces. While significant progress has been made in developing autonomous agents that operate in isolation, deploying them in dynamic, multi-agent environments presents new theoretical, algorithmic, and societal challenges. Towards this goal, I will introduce a novel theoretical framework—Markov near-potential functions (MNPFs)—to study multi-agent interactions in dynamic environments. I will demonstrate how this framework can be leveraged to design competitive, real-time planning and control strategies for autonomous multi-car racing. I will also present a new multi-agent reinforcement learning pipeline – Near-Potential Policy Optimization – which exploits the structure of MNPFs to compute low-regret approximate Nash strategies in general-sum dynamic games.

Human-interactive driving can be viewed as a collaborative game between the vehicle’s AI and the human driver. However, several factors, such as unobservable short- and long-term driver-state factors and variations in the reward models, make human-interactive driving more challenging as an ML problem, leading to a diverse set of approaches to assist the driver. In this talk, I will survey how my team addresses some of these challenges, spanning a full pipeline of assistance: from understanding driver state (gaze, situational awareness, cognitive factors), to predicting driver behavior and interacting with drivers across different time scales. On the interaction side, we demonstrate the use of fictitious play in a world model for RL driver assistance, imitation of human coaches, and learning-aware assistance policies that explicitly target skill development rather than task performance alone -- reducing skill atrophy while guiding drivers toward safer, more capable behavior.