An arXiv preprint — an open, non peer reviewed repository — from IIMB Bangalore (the Indian Institute of Management Bangalore) and University of Antwerp researchers swaps fixed recharge rules for a fleet wide learned policy that picks both where
A simulated warehouse fleet learned to choose its own charging station and plug-in duration, beating the strongest hand-coded baseline by roughly 6% in order-completion rate (arXiv:2607.05683). That modest gap comes from a single architectural move: a static per-robot battery threshold replaced by a coordinated, fleet-wide learned policy.
Sobhanan (IIMB Bangalore) and Christof Defryn (University of Antwerp) trained autonomous mobile robots (AMRs, the wheeled pickers that move inventory in modern warehouses) with Proximal Policy Optimization, a reinforcement-learning method where the policy is shaped by trial-and-error against the simulator's rewards (University of Antwerp faculty page). Each robot's learned action is a single joint decision: which charging station to walk to and how long to stay, with explicit anticipation of queue time at that station. Picking station and plug-in duration jointly, rather than as separate steps, keeps the two decisions from working against each other.
The result is in a simulated multi-block warehouse with fixed chargers and stochastic order arrivals. It is the authors' own benchmark, not a deployed rerun on a real customer fleet, and arXiv preprints are not peer-reviewed. The mechanism travels; the 6% is the simulation's evidence that the new shape works.