Publications

Recent advances in multi-population mean-field games (MP-MFGs) typically rely on expected-value objectives and do not account for uncertainty in the behavior of other populations. We introduce risk-averse MP-MFGs, a framework in which each population optimizes its worst-case expected reward over dynamically feasible ambiguity sets centered on the other team's nominal mean-field flows. We develop a risk-averse fictitious-play algorithm and show, under mild monotonicity and regularity conditions, that risk-averse exploitability serves as a Lyapunov function and decays at a rate O(1/s), despite the additional nonlinearity introduced by the worst-case objective. Numerical experiments demonstrate convergence and illustrate the resulting risk-averse behaviors in multi-population decision-making.

While MARL has shown strong empirical performance, existing methods struggle with a large number of agents due to the combinatorial growth of joint interactions. Mean-field (MF) approximations address this challenge by replacing pairwise interactions with surrogate interactions against population distributions, yielding tractable formulations for solving large-population policies. However, existing MF formulations focus on fully cooperative or purely competitive settings and fail to capture the mixed cooperative–competitive structure of team-based games. We extend mean-field learning to zero-sum team games, where agents cooperate within each team and compete at the team level. We show that such games admit ε-optimal decentralized policies that depend only on each agent's local state and population team distributions. Building on this structure, we propose MF-MAPPO, a scalable algorithm with a shared actor and a minimally informed critic per team. MF-MAPPO is trained directly in finite-population simulators rather than using mean-field oracles, enabling deployment to realistic scenarios with thousands of agents. We further extend MF-MAPPO to partially observable settings via a simple gradient-regularized training scheme. Experiments on large-scale benchmarks in our simulation platform (MFEnv), including population games with analytical solutions and high-dimensional battlefield scenarios, demonstrate that MF-MAPPO outperforms existing MARL baselines and yields rich, heterogeneous behaviors.

Mean-field games address scalability in large-population multi-agent systems by reducing inter-agent interactions to those between a representative agent and an aggregate population distribution. However, standard mean-field problem formulations assume either fully rational agents and/or a fixed and known initial agent distribution. To address these limitations, we introduce a unified framework that incorporates risk aversion to uncertainty in the initial population distribution together with an explicit quantal-response model of bounded rationality, leading to a generalized equilibrium concept, which we call the mean-field risk-averse quantal response equilibrium (MF-RQE). We prove the existence of MF-RQE, establish convergence guarantees for fixed-point iteration and fictitious play, and propose a scalable reinforcement learning algorithm for large state–action spaces. Numerical experiments show that MF-RQE yields superior robustness over classical MFG baselines under distributional shifts and agent suboptimality.

In this paper, we explain how differential flatness combined with Gaussian Processes can be applied for the adaptive cruise control system under external, unknown dynamics, to attain accurate trajectory tracking. The proposed approach employs differential flatness for controller design, along with Gaussian Processes to estimate the unknown components in the dynamics. In order to ensure robustness to this data-driven technique, a robust control law is also introduced, that further exploits the Gaussian Process framework. In the recent publication, we showed the effectiveness of the proposed control methodology for set-point regulation. In the current paper, we generalize it to trajectory tracking and show numerically that the error dynamics converge to zero.

This paper develops a novel control methodology for the adaptive cruise control system under external, unknown dynamics to attain set point regulation. The approach employs differential flatness for controller design, along with Gaussian Processes to estimate the unknown components in the dynamics. In order to ensure robustness to this data-driven technique, a robust control law is also introduced, that further exploits the Gaussian Process framework. The error dynamics is shown to converge to zero with high probability. Numerical experiments to support the results are presented.

This paper presents a methodology to control a spherical pendulum on a stationary base in an upright position. This is based on the application of a nonlinear PD controller and a linear feedback controller in a small domain around the vertically upward equilibrium. The controller is designed to transition from the nonlinear to the linear region once it becomes sufficiently close to its equilibrium point. The region of asymptotic stability of the pendulum encompasses all initial conditions apart from the vertically downward position. Numerical experiments illustrating the performance of the controller are also provided.

In this paper, prospects of the utilization of the 4th stage of ISRO’s PSLV, after the completion of the launch mission, as an orbital platform, to host scientific payloads have been discussed. Payloads from 4 domains - Technological Demonstration, Earth Observation, Microgravity, Biology Experiments and the associated mission concepts have been surveyed, and comments have been made on their suitability to be launched onboard the orbital platform. Technological challenges in achieving these have been highlighted. Based on this analysis, two technology demonstration missions have been proposed by the team.