When the Warehouse Burned, the Robots Kept Working

The alarm sounded at 2:30 a.m. on February 5, 2019, at Ocado's Andover fulfillment center in Hampshire, England. A fire had broken out in the facility's battery charging area, quickly spreading through the three-story grid where hundreds of robots whisked grocery crates between storage and packing stations. The blaze would eventually require 200 firefighters and burn for three days, gutting the building and causing £100 million in damage. But in the first critical minutes, something remarkable happened: the warehouse's brain — a multi-agent coordination system governing more than 1,000 robots — reacted faster than any human operator could.

The grid goes dark

Ocado's "hive" is a dense lattice of aluminum tracks covering an area the size of several football fields. At any given moment, hundreds of box-shaped robots — each about the size of a washing machine — zip across the grid at up to four meters per second, lifting and lowering crates of groceries through vertical chutes to human pickers below. The robots don't follow pre-programmed paths. Instead, each one is an autonomous agent that continuously negotiates its route with a central traffic controller and with neighboring robots, optimizing for speed, battery life, and collision avoidance.

When the fire started, the system's first response wasn't to shout "fire." It was to notice that a cluster of robots near the charging stations had stopped reporting their positions. The central coordinator, which maintains a real-time model of every robot's location, velocity, and task queue, detected the anomaly within seconds. It didn't wait for a human to confirm the emergency. It immediately reclassified that sector of the grid as impassable and began recalculating paths for every other robot in the fleet.

Rerouting in real time

The coordination engine uses a variant of the Contract Net Protocol, a market-based mechanism where tasks — "move crate X to station Y" — are auctioned to the best-positioned robots. When a section of the grid goes dark, the controller doesn't just pause those tasks; it instantly re-announces them to robots elsewhere in the hive. Within milliseconds, new bids come in, new routes are assigned, and the swarm reorganizes around the hole in the grid.

In the Andover fire, this rerouting happened continuously as the blaze spread. Robots that had been heading toward the affected zone peeled off and joined new task queues. Robots with low battery that would normally have returned to the now-burning chargers were redirected to standby chargers on the opposite side of the facility. The system even adjusted the pickers' work assignments, routing crates to stations that remained reachable.

No human dispatcher typed a single reroute command. The multi-agent system — built on principles of decentralized decision-making, stigmergic coordination through shared state updates, and real-time auction-based task allocation — executed thousands of micro-decisions per second, keeping the unaffected 80% of the warehouse operational while firefighters battled the flames above.

Designing for the unexpected

This resilience wasn't accidental. Ocado's engineers had spent years simulating catastrophic failures in digital twins of the hive — virtual replicas where they could inject faults, network partitions, and yes, simulated fires, to see how the agent ecosystem would respond. The system's architecture drew explicitly on research into Byzantine fault tolerance: the ability to maintain correct operation even when some components behave arbitrarily (or maliciously). In this case, the "faulty" components were robots that had gone silent because their comms melted, and a chunk of the environment that had become physically impassable.

The key insight, borrowed from both robotics and distributed systems theory, was that coordination shouldn't rely on a single central planner. While a central traffic controller exists, the robots themselves maintain local situational awareness and can execute collision-avoidance maneuvers even if the central link drops. The system degrades gracefully: lose the central coordinator, and the robots fall back to local rules; lose a sector, and the rest of the swarm rebalances.

Aftermath and implications

The Andover fire destroyed the building and the robot fleet inside it. Ocado's stock price dropped 13% the next day. But the company's other fulfillment centers — running the same multi-agent software — continued operating without interruption. Within months, a new, larger hive was operational at their Erith site, incorporating lessons from the fire: better thermal segmentation, more distributed charging infrastructure, and enhanced smoke detection integrated directly into the agent coordination loop.

The incident became a landmark case study in the multi-agent systems community. It demonstrated that the theoretical promises of decentralized coordination — robustness, scalability, graceful degradation — hold up under real-world catastrophe. The robots didn't panic. They didn't freeze. They did what they were designed to do: perceive, communicate, negotiate, and adapt, thousands of times per second, keeping the grocery moving while the building burned around them.

This is one episode in a much longer story. For the full account of multi-agent systems in logistics, read “Multi-Agent Systems and Coordination” by Maria Meyer on MixCache.com.