We develop fleet management, task allocation, and multi-agent pathfinding technologies that coordinate
large teams of robots efficiently and reliably in shared, dynamic environments.
Robot Fleet Management System (FMS)
Real-time traffic control: real-time optimal path planning and traffic control that prevents collisions and deadlocks among multiple AMRs on mixed human-robot smart-manufacturing floors.
Large-scale fleet operation: combines decentralized control (flexibility) with centralized control (efficiency) to operate fleets of 100+ robots stably and without delay.
2 SCIE journal papers · 5 technology transfers
A semi-centralized robot fleet management system (FMS) for logistics automation.
Intermittently provides sector-based global information to a decentralized, learned MAPF policy, approaching the performance of centralized planners.
Combines imitation-learning-based local planning with intermittent sector-congestion guidance.
An imitation-learning-based decentralized multi-agent pathfinding (MAPF) algorithm.
Multi-Agent Pickup and Delivery
A TSP-based optimal task-allocation algorithm for multi-agent pickup and delivery.
Task group allocation: efficiently assigns real-time incoming orders by considering each robot's route and expected completion time.
Task grouping: groups tasks by each robot's capacity and expected cost before assignment, for efficient throughput.
Task group updates: continuously compares unassigned and pending tasks to keep task groups efficient at all times.
1 SCIE journal paper
TSP-Based Task Allocation for Multi-Task Multi-Agent Pickup and Delivery — ICROS 2024
Efficient TSP-Based Task Group Allocation for Multi-Task Multi-Agent Pickup and Delivery — IEEE Access, 2025
Multi-Agentic AI
Hierarchical agent collaboration: a central agent handles global planning while per-group local agents own detailed scheduling and re-planning, reducing LLM context load and mitigating hallucination.
Localized failure recovery: on execution failure, only the affected group pauses to re-plan while the rest of the fleet keeps executing in parallel, uninterrupted.
Failure-injection benchmark: introduces RECAP-THOR, a new dataset for controlled evaluation of re-planning capability.
A hierarchical LLM-agent-based task planning and scheduling system for long-horizon multi-robot missions.
