Aditya Kapoor

The development of embodied agents that can effectively interact in the physical world represents a significant challenge, particularly when extended to multi-agent systems where coordination, collaboration, and competition are critical. This project focuses on designing and developing intelligent, embodied agents using deep reinforcement learning (DRL) and generative models to enable robust, real-time decision-making in dynamic and uncertain environments.

Key to this research is the exploration of multi-agent systems, where agents must interact not only with their environment but also with other agents. This presents challenges such as learning optimal policies in cooperative or competitive scenarios and ensuring effective communication and coordination among agents. The use of DRL will enable agents to learn these complex, long-horizon tasks in an interactive manner, while generative models will allow them to simulate and predict potential interactions with the environment and other agents, providing more efficient and adaptable decision-making strategies.

By focusing on actor-critic methods that guarantee convergence to optimal policies, this project will explore novel architectures for multi-agent reinforcement learning, addressing issues such as scalability, partial observability, and emergent behaviors in agent interactions. The overarching goal is to develop systems where agents can seamlessly collaborate or compete in real-world tasks, adapt to new environments, and generalize their learning to novel, unseen settings.