Addison Kalanther

I’m a first-year EECS PhD student at UC Berkeley, advised by Professor Shankar Sastry in Berkeley Artificial Intelligence Research (BAIR). I study reinforcement learning, robot learning, and multi-agent learning, broadly aiming to build machines that reason, with a focus on robotics. I also work on autonomous manufacturing at Tensr, edit the BAIR blog, and manage the BAIR website.

Research

Orchestrating Level-K Policies Against Unknown Opponents in Partially-Observable Dynamic Games

A. Kalanther, S. Bharvirkar, D. Bostwick, C. Maheshwari, S. Sastry
Under review, 2026
arXiv
We study how to deploy pretrained level-K policies against an opponent whose reasoning level is unknown. In pursuit-evasion, an orchestrator trained with RL to choose among them earns higher return than classifiers that estimate the opponent's level and play the matching response.

NePPO: Near-Potential Policy Optimization for General-Sum Multi-Agent Reinforcement Learning

A. Kalanther, S. Bharvirkar, S. Sastry, C. Maheshwari
IEEE Conference on Decision and Control (CDC), 2026 [Accepted]
project page / arXiv
We propose NePPO, a multi-agent RL algorithm for approximating Nash equilibria in general-sum games. It learns a single potential function shared by all players, so that an equilibrium of the cooperative game it defines approximates one of the original game.

Coordinated Autonomous Drones for Human-Centered Fire Evacuation in Partially Observable Urban Environments

M. Mendoza, A. Kalanther, D. Bostwick, E. Stephan, C. Maheshwari, S. Sastry
IEEE Global Humanitarian Technology Conference (GHTC), 2025
project page / arXiv
We train heterogeneous UAVs to guide human evacuees through a simulated urban fire, modeling it as a POMDP and using RL to learn policies robust to unknown panic behavior.

Evader-Agnostic Team-Based Pursuit Strategies in Partially-Observable Environments

A. Kalanther, D. Bostwick, C. Maheshwari, S. Sastry
Workshop on Multi-Agent Systems, Robotics: Science and Systems (RSS), 2025
project page / arXiv
We train a team of two pursuer UAVs to intercept an evader of unknown behavior in an occluded urban environment. Deep RL builds a level-K hierarchy of policies offline; online, a classifier estimates the evader's level and plays the matching response.