University of Massachusetts Amherst
I study how robots learn to work together.
Human-Centered Robotics Lab · Working with Hao Zhang
Safety and fairness restrict how a robot team moves without rewarding task completion, so a single scalar reward has to fix their trade-off in advance. SF-Push keeps one policy gradient per robot and per objective, takes its direction from a scale-free minimum-norm combination and its step length from the task gradient alone — no trade-off weights to tune.
Rewarding task progress alone makes a quadruped pair either crowd into unsafe spacing or leave one robot idle. Adding team-level proximity and fairness shaping to MAPPO fixes both.
Composing new objectives into an already-trained multi-robot policy at execution time, without retraining or tuning trade-off weights.
Grounding an abstract simulator with real-world data and state history so policies trained in it transfer to physical robots.
Workshop version of the flow-policy guidance work. Presented at the ICRA 2026 Workshop on Multi-Agent Robotic Systems and the RSS 2026 Workshop on Diffusion for Robot Learning.
Combining learned behaviors and a classical robotics stack for autonomous team play. The system won the 2024 RoboCup SPL Challenge Shield Division.
How can robots coordinate their actions while balancing shared goals and individual objectives?
MULTI-AGENT REINFORCEMENT LEARNINGHow can policies learned in simplified simulations remain useful when deployed on physical robots?
ABSTRACTION & POLICY TRANSFERPh.D. student, Computer Science
B.S., Computer Science
B.E., Electrical & Computer Engineering
UM–SJTU Joint Institute
University of Massachusetts Amherst
Research with Hao Zhang
University of Wisconsin–Madison
Research with Josiah P. Hanna and Peter Stone