PPO-based DRL boosts warehouse order rates by 6% while cutting recharge time

A new paper from arXiv (2607.05683) tackles a critical operational bottleneck in warehouse automation: battery management for Autonomous Mobile Robots (AMRs). Traditional fixed-rule heuristics and simplistic charging policies often fail under stochastic order arrivals and multi-robot coordination, leading to downtime and throughput losses. To address this, researchers Shaji, Sobhanan, and Defryn propose a deep reinforcement learning framework based on Proximal Policy Optimization (PPO). The model learns two decisions simultaneously: which charging station to use (among fixed stations in a multi-block warehouse) and how long to charge, explicitly anticipating queuing times at stations. This dynamic policy balances the trade-off between keeping robots available for order picking and ensuring they have enough charge to complete tasks without interruption.

Extensive numerical experiments benchmarked the PPO approach against state-of-the-art DRL methods and traditional heuristics across varied warehouse layouts and arrival rates. Results show the PPO framework achieves up to a 6% increase in order-completion rates compared to the strongest baseline, while also cutting the total time robots spend recharging—a dual benefit that directly improves warehouse throughput and energy efficiency. The learned policy was interpretable, offering insights into when and where robots should charge to minimize system-level delays. The work is particularly relevant as warehouses scale up AMR fleets and face dynamic demand patterns, positioning DRL as a practical alternative to hand-coded rules for real-time operational decisions.