Search & Pursuit
Adversarial Multi-Agent Pursuit-Evasion in Complex Stochastic Environments
This project develops a decentralized, learning-based framework for visibility-based pursuit-evasion in challenging outdoor environments. We focus on enabling teams of mobile agents to systematically clear contaminated spaces and capture adversarial evaders within high-density urban terrains. By integrating multi-agent reinforcement learning with advanced spatial reasoning, the system addresses the critical challenges of building-induced occlusions and limited sensor ranges, allowing for real-time coordinated maneuvers without the need for a central controller.
Multi-Agent RL (MARL)Multi-Agent CoordinationGraph Attention NetworksDecentralized Control
Contributions
What this work contributed, drawn from the papers below.
-
01
Guaranteed detection under limited sensing
A team of pursuer robots with limited sensing must detect every evader in a multiply-connected planar environment, where evaders are omniscient and potentially arbitrarily fast and the map may not be known in advance. This demands tight coordination to guarantee detection.
ViPER, CoRL 2024 -
02
A distributed policy from graph attention
ViPER uses a graph attention network to learn a coordinated yet distributed policy through multi-agent reinforcement learning, instead of the large teams that existing non-learning methods need to clear an environment.
ViPER, CoRL 2024 -
03
Outperforms non-learning planners
Experiments show ViPER significantly outperforming state-of-the-art non-learning planners on the same pursuit-evasion tasks.
ViPER, CoRL 2024 -
04
Emergent coordination that scales with team size
Coordinated behaviours emerge from training and carry over to more challenging scenarios and to varying team sizes.
ViPER, CoRL 2024 -
05
Deployed on hardware
The learned policies were deployed on physical robots in an aerial search task.
ViPER, CoRL 2024