Multi-Agent Path Planning for Warehouse Robotics

MSc Thesis - Technion, Israel Institute of Technology

Problem

Modern warehouse automation requires multiple robots to navigate shared spaces efficiently while avoiding collisions. The challenge: coordinate dozens of agents in real-time, balancing throughput against safety and computational constraints.

Why It's Hard

• Multi-agent coupling: Each robot's decision affects all others. The state space grows exponentially with the number of agents.
• Partial observability: Robots have limited sensing range and cannot predict other agents' intentions perfectly.
• Real-time constraints: Decisions must be made in milliseconds to maintain smooth operation.
• Sim-to-real gap: Algorithms that work in simulation must transfer to physical robots with sensor noise, latency, and mechanical imperfections.

Approach

Investigating coordination strategies that balance optimality with computational tractability:

• Planning algorithms: Comparing centralized vs decentralized approaches (CBS, prioritized planning, MAPF variants)
• Social laws: Implicit coordination rules that reduce conflicts without explicit communication
• Collision avoidance: Integrating safety constraints into the planning loop

Current Milestones

Completed:

• Literature review of MAPF algorithms and multi-agent coordination
• Simulation environment setup (highway-env, custom grid worlds)
• Baseline implementation of standard planning algorithms

In Progress:

• Designing hybrid coordination strategies
• Benchmarking on warehouse-scale scenarios

Next:

• Evaluation on realistic warehouse layouts
• Analysis of scalability and failure modes

Evaluation Metrics

• Success rate: Percentage of agents reaching goals without collision
• Makespan: Time for all agents to complete their tasks
• Throughput: Tasks completed per time unit
• Collision rate: Safety violations during execution
• Computation time: Planning overhead per decision cycle

Technologies

Python, Multi-Agent Reinforcement Learning, Graph-based Planning (A*, RRT*), Simulation Environments (highway-env, Isaac Sim)

Media

Simulation visualizations and results coming soon