Multi-Robot Warehouse System Swarm Intelligence

Project Context

The client had recently expanded from a single-shift to a 24/7 operation to meet growing order volumes, but their existing fleet of manually dispatched AGVs could not keep pace. Peak-season congestion was causing missed SLAs, and the facility layout — a converted retail warehouse with narrow aisles and mixed mezzanine levels — made off-the-shelf fleet management solutions impractical without significant customisation.

Project Overview

Developed a scalable fleet management system for a 120,000 sq ft e-commerce fulfilment centre near Leeds. The system orchestrates the movement of up to 50 AMRs, optimising traffic flow, preventing deadlocks, and ensuring efficient task completion in a high-density environment with narrow aisles and mixed mezzanine levels.

The Challenge

The client needed to scale from 10 manually dispatched AGVs to a coordinated fleet of 50 AMRs within their existing 120,000 sq ft facility. Bottlenecks with traffic congestion and inefficient robot coordination were causing missed delivery SLAs during peak periods. Key challenges included:

• Coordinating up to 50 robots in a shared, dynamic space with narrow aisles
• Preventing deadlocks in single-lane corridors and intersection bottlenecks
• Optimising battery charging schedules to maintain throughput across 24/7 operations
• Integrating robots from different vendors (heterogeneous fleet of 3 models)
• Ensuring low-latency communication for real-time control across the full facility footprint
• Dynamic task reallocation based on priority and robot status

What Had Been Tried Before

Prior to engaging YF Studio, the client had attempted two approaches. First, they trialled the built-in fleet manager provided by their primary robot vendor, but it could not coordinate the two additional robot models in their fleet and had no deadlock recovery mechanism — robots would simply stop and wait for manual intervention. Second, they contracted a systems integrator to build a centralised dispatcher, but it treated path planning as a static problem: routes were precomputed at shift start and could not adapt to real-time congestion, leading to cascading delays whenever a single robot deviated from its plan. Both solutions failed to scale beyond approximately 15 robots before congestion became unmanageable.

Our Solution

1. Centralised Traffic Manager

Implemented a global traffic management system using Multi-Agent Path Finding (MAPF) algorithms, specifically the Conflict-Based Search (CBS) variant. This system reserves space-time corridors for each robot, guaranteeing collision-free paths and preventing deadlocks.

Why CBS over alternatives: We evaluated both Priority-Based Search (PBS) and decentralised ORCA-style velocity obstacles. PBS was faster to compute but produced suboptimal paths in the client's narrow-aisle layout, increasing average travel time by 18%. ORCA could not guarantee deadlock-freedom in single-lane corridors. CBS provided provably optimal, collision-free plans within acceptable compute time (<200ms for 50 robots) when combined with our hierarchical zone decomposition.

2. Heterogeneous Fleet Adapter

Developed a standardised interface layer using Open-RMF (Robotics Middleware Framework) to allow robots from three different manufacturers to communicate and coordinate within the same system.

Why Open-RMF over custom middleware: The client's fleet included robots from three vendors, each with proprietary APIs. Building custom adapters for each would have tripled integration effort and created an ongoing maintenance burden. Open-RMF provided a proven abstraction layer with existing fleet adapter templates, reducing integration time from an estimated 3 months to 6 weeks per vendor.

3. Dynamic Task Auctioning

Created a market-based task allocation system where robots "bid" on tasks based on their location, battery level, and current capabilities, ensuring optimal resource utilisation.

Why auction-based over greedy assignment: The client's previous system used nearest-available assignment, which led to fleet imbalance — robots near high-demand zones were over-utilised while others sat idle. The auction mechanism naturally distributes load by factoring in battery state, travel distance, and current congestion, yielding 22% better fleet utilisation in our simulation benchmarks compared to greedy dispatch.

4. Predictive Maintenance & Charging

Integrated a predictive model that monitors battery health and operational patterns to schedule charging and maintenance during low-demand periods, maximising fleet uptime.

Why predictive over threshold-based charging: Simple low-battery thresholds caused charging station congestion during shift changes when multiple robots hit the threshold simultaneously. Our predictive model staggers charging across demand troughs, reducing peak charging station occupancy by 40% and eliminating the "charging rush" that previously took up to 8 robots offline simultaneously.

Technical Implementation

System Architecture

The solution is built on a layered architecture:

• Fleet Level: Global route planning and task dispatching (Open-RMF)
• Robot Level: Local navigation and obstacle avoidance (ROS2 Nav2)
• Network Level: Low-latency state synchronization via Zenoh/DDS
• Interface Level: WMS integration via REST/gRPC APIs

Conflict Resolution

Advanced deadlock prevention strategies:

• Priority-based Yielding: Lower priority robots yield to high priority tasks
• Replanning: Dynamic rerouting around congested areas
• Traffic Lights: Virtual traffic control at high-density intersections
• Buffer Zones: Dedicated waiting areas for idle robots

Limitations & Edge Cases

• Maximum Fleet Size: The system has been tested with up to 50 robots in a single facility. Beyond this, the CBS planner's computation time exceeds the 200ms real-time constraint without further architectural changes (e.g., sharding the planner across multiple zones).
• Cross-Zone Handoffs: When robots transition between coordination zones, handoff latency adds approximately 200ms. For time-critical tasks spanning multiple zones, this must be accounted for in path cost estimates.
• Network Dependency: The system requires a dedicated 5 GHz Wi-Fi network. On shared networks, packet loss above 20 robots causes unacceptable command latency spikes (>100ms), leading to precautionary stops and throughput degradation.
• Heterogeneous Speed Limits: Robots with significantly different maximum speeds (e.g., forklifts vs. shelf-movers) create convoy effects in single-lane corridors. We mitigate this with speed-class-aware routing, but it does not fully eliminate the issue in heavily constrained layouts.
• Manual Intervention Recovery: If a human manually moves a robot during operation (e.g., during maintenance), the system requires a re-localisation cycle of 5–15 seconds before the robot can rejoin the coordinated fleet.

Results & Impact

Operational Metrics (Before / After)

• Average robot utilisation: 60% → 85%
• Average task completion time: reduced by 25% (from 4.2 min to 3.1 min per pick cycle)
• Main corridor congestion incidents per shift: 12–15 → fewer than 1
• Fleet integration: single vendor → 3 robot models coordinated under one system
• Deadlock events requiring manual intervention: ~8 per week → avg. 1 recoverable deadlock per 6 weeks (auto-resolved within 10 seconds)
• Scaling from 10 to 50 robots required a Wi-Fi 6 infrastructure upgrade at 30 units to maintain sub-50ms command latency. Above 35 robots, we switched from centralised to hierarchical zone-based coordination to keep the MAPF planner within its real-time compute budget.

System Capabilities

The system manages various warehouse operations:

• Goods-to-Person: Moving shelves to picking stations
• Pallet Transport: Heavy-payload autonomous forklifts
• Conveyor Bridging: Transferring items between conveyor belts
• Trash Removal: Autonomous waste bin collection
• Inventory Scanning: Mobile scanning robots for stock taking

Safety & Reliability

Ensuring safe operation in human-robot environments:

• Safety Zones: Dynamic speed limiting near human workers
• Redundancy: Fail-over servers for fleet management
• Emergency Stop: Global e-stop capability for the entire fleet
• Simulation Testing: 1000+ hours of simulation before deployment
• Cybersecurity: Encrypted communication and strict access control

Integration

Deep integration with enterprise systems:

• WMS Integration: Real-time order ingestion from SAP/Oracle
• Facility Integration: Control of automatic doors and elevators
• Dashboard: Real-time 3D visualization of fleet status
• Analytics: Historical data analysis for process optimization

Ongoing & Next Steps

Active development and planned near-term improvements:

• In progress: Digital Twin integration using Isaac Sim for offline route optimisation and new-layout validation before physical reconfiguration
• In progress: Reinforcement learning for decentralised collision avoidance at intersections, aiming to reduce the CBS planner's computational load by handling local conflicts on-robot
• Planned (Q3 2024): Tablet-based interface for floor supervisors to manually re-prioritise tasks and temporarily block zones during restocking
• Under evaluation: Integration with outdoor AMRs for yard-to-dock goods transfer, pending weatherproofing and GPS-to-indoor-localisation handoff testing
• Under evaluation: Adding mobile manipulators for collaborative pick-and-pack, which would require extending the MAPF planner to account for arm reach envelopes in collision checking