Every day, hospitals orchestrate thousands of silent deliveries — medications, specimens, supplies — each carrying its own urgency and its own deadline. Researchers have now built an artificial intelligence system called HMART that teaches robot fleets to understand the difference between what can wait and what cannot, reducing average delivery delays by over 98 percent in simulation. The work sits at the intersection of logistics and care: a reminder that the unglamorous movement of things through hallways is, in medicine, a matter of consequence. Whether the gains survive contact with real b
AI-powered robot scheduling cuts hospital delivery delays by 98%
Hospitals move thousands of items every day. Each has a deadline.
Mark
So this is about robots moving things around hospitals faster. Why does that matter enough to publish in Nature?
Mimi
It matters because hospitals move thousands of items daily, each with different time windows. A stat medication might need to arrive in fifteen minutes. A routine lab specimen can wait two hours. The system had to balance these competing demands while coordinating multiple robots through constrained spaces—elevators, hallways. That's a hard optimization problem.
Luke
But this is all simulation, right? Four floors, three elevators, twelve robots. That's not a real hospital.
Mimi
Correct. It's a discrete-event simulator. The team is explicit about that. They say the gains require confirmation in a live building.
Mark
What does the improvement actually look like in numbers?
Mimi
Average delivery delay dropped from 2,359 minutes to 29 minutes. On-time delivery for urgent medications went from 92.5 percent to 96.9 percent. The worst-case delays fell by more than a factor of ten.
Luke
Those are big numbers. But I want to know: is 2,359 minutes what hospitals are actually experiencing now, or is that what the baseline algorithm produces in simulation?
Mimi
That's the baseline algorithm in simulation. The current dispatching rules hospitals use. So we're comparing HMART against the strongest existing algorithm, not against what hospitals actually do today.
Mark
Can the system handle different hospital layouts?
Mimi
The trained version, HMART-RL, transferred to configurations it had never seen during training. It also scaled across different fleet sizes without retraining. That's significant because it means you would not need to retrain the system for every hospital.
Luke
How fast does it make decisions?
Mimi
Well under a second per epoch. That matters because hospitals cannot afford to wait for the computer to think.
Mark
What's the catch?
Luke
The catch is that this is simulation. Real robots break down. Real hallways have people in them. Real elevators have actual wait times that might not match the model. The team knows this and says so. But we do not yet know if a 98 percent improvement in simulation translates to a meaningful improvement in a real hospital.
Mimi
That's fair. But the fact that they published the code and data for independent verification, and that they were explicit about the limitations, suggests they are serious about moving toward real-world testing.
Mark
So what happens next?
Mimi
Someone needs to deploy this in an actual hospital and see if it works.
O Pulso
- Hospital robot fleets face a genuinely hard problem — dozens of competing requests, shared elevators, and clinical deadlines that range from minutes to hours, all demanding resolution in real time.
- Without intelligent coordination, delays compound: simulated baseline conditions showed 80 requests waiting at any moment and average tardiness stretching past 2,300 minutes.
- HMART attacks the problem hierarchically — assigning requests by priority, optimizing routes, and managing elevator access through a model predictive control layer that prevents robots from bottlenecking at the lift.
- The results in simulation are striking: mean tardiness collapsed from 2,359 minutes to 29, and on-time delivery of critical medications climbed from 92.5 to 96.9 percent.
- The trained system transfers to hospital configurations it has never seen, scaling across fleet sizes without retraining — but real-world deployment remains the uncleared threshold.
Every day, hospitals orchestrate thousands of silent deliveries — medications, specimens, supplies — each carrying its own urgency and its own deadline. Researchers have now built an artificial intelligence system called HMART that teaches robot fleets to understand the difference between what can wait and what cannot, reducing average delivery delays by over 98 percent in simulation. The work sits at the intersection of logistics and care: a reminder that the unglamorous movement of things through hallways is, in medicine, a matter of consequence. Whether the gains survive contact with real buildings and real clinical chaos is the question that remains.
Hospitals move thousands of items every day, and each one has a deadline. Some are urgent. Some can wait. The robots tasked with these deliveries must constantly decide which request comes first, which route to take, and how to share elevators without creating gridlock — a problem that looks manageable until you try to solve it at scale.
Researchers built HMART, a hierarchical multi-agent reinforcement learning system, to manage exactly these decisions. It assigns requests to robots in real time, weighing the difference between a stat medication that must arrive within fifteen minutes and a routine specimen that can wait two hours. Tested in a detailed simulation of a four-floor hospital running twelve robots under heavy load, the system was compared against the best existing dispatching algorithms currently used in practice.
The gap was not subtle. Average delivery delay fell from 2,359 minutes to 29. Pending requests waiting for pickup dropped from 80 to 14. On-time delivery for critical medications improved from 92.5 to 96.9 percent, and worst-case delays shrank by more than a factor of ten. The system makes its decisions in well under a second — a practical necessity, since hospitals cannot pause for computation.
HMART achieves this by treating the hospital not as a collection of independent robots but as a coordinated system. A shared encoder gives each agent awareness of the building's layout and current state. A risk constraint prevents the system from trading catastrophic individual delays for a better average. Elevator access is managed through a predictive control layer that keeps robots from queuing at the same lift simultaneously. The trained version of the system transferred to hospital sizes and configurations it had never encountered during training.
The researchers are careful about what they have and have not shown. Every result comes from simulation. The code and data are published for independent verification, and the team is explicit that confirmation in a real building — with actual robots, actual elevators, and actual clinical staff — is the necessary next step. A 98 percent reduction in average delay is not a marginal finding, and a four-point improvement in stat medication delivery could translate directly to patients waiting less for pain relief or antibiotics. But the distance between a well-specified simulation and a working hospital remains the open question.
Hospitals move thousands of items every day—blood samples to the lab, medications to the ward, supplies to surgery. Each delivery has a deadline. Some are urgent. Some can wait. The robots that carry these items must decide which request to handle first, which route to take, and how to share the elevators without creating bottlenecks. It's a problem that looks simple until you try to solve it.
Researchers at a major institution built a scheduling system called HMART—hierarchical multi-agent reinforcement learning for transport—to manage these decisions. The system uses artificial intelligence to assign requests to robots and optimize their routes in real time, accounting for the fact that different types of deliveries have different time windows. A stat medication order might need to arrive within fifteen minutes. A routine lab specimen can wait two hours. The system weighs these competing demands and coordinates the movement of multiple robots through a constrained space.
The team tested HMART in a detailed computer simulation of a four-floor hospital with three elevators and twelve robots operating under heavy load. They compared it against the strongest existing dispatching rules—the algorithms hospitals currently use to send robots on their routes. The results were stark. The system reduced average delivery delay from 2,359 minutes to 29 minutes. Work-in-process inventory—the number of requests waiting to be picked up—dropped from 80 to 14. Most clinically significant, the on-time delivery rate for stat medications improved from 92.5 percent to 96.9 percent. The ninety-fifth percentile tardiness—a measure of how bad delays get at their worst—fell by more than a factor of ten.
The system comes in two versions. HMART-H uses hand-engineered decision rules. HMART-RL trains a neural network to make assignment decisions while using the engineered rules for route optimization and elevator coordination. Both versions run in near-real time, making decisions in well under a second per epoch, which matters because hospitals cannot afford to wait for a computer to think. The trained version, HMART-RL, proved transferable: it worked on hospital configurations it had never seen during training, and it scaled across different fleet sizes without retraining.
The gains emerged from how the system frames the problem. Instead of treating each robot as an independent agent, HMART models the hospital as a multi-objective game where robots coordinate through a shared encoder that understands the building's layout and current state. A conditional value-at-risk constraint ensures that the system does not sacrifice worst-case performance to optimize the average—it bounds the tail of delays, protecting against the scenario where a few requests get catastrophically delayed. A model predictive control layer manages elevator access, preventing the bottleneck that occurs when multiple robots queue for the same lift.
The research is thorough about its limitations. All results come from simulation. The hospital tested was four floors with thirty-two service points and three elevators. The demand mix, deadlines, and robot fleet size were all specified. The team ran twenty paired random seeds for each experiment to ensure results were not artifacts of a particular scenario. They published their code and data to allow independent verification. But they are explicit: these gains require confirmation in a real building with actual robots, actual elevators, and actual clinical staff.
What matters here is not the technical sophistication—though the use of graph attention networks and hierarchical reinforcement learning is substantial—but the scale of the improvement and its specificity to clinical constraints. A 98 percent reduction in average delay is not marginal. A four-percentage-point improvement in stat medication on-time delivery could mean fewer patients waiting for pain medication or antibiotics. The system works because it understands that not all delays are equal. It is built to protect the urgent while efficiently handling the routine. Whether those gains hold when robots collide with real hallways, real staff, and real clinical chaos remains to be seen.
Citações Notáveis
These gains are obtained in simulation and require confirmation in a live building.— Research team, Nature publication
Quer a matéria completa?
Leia o original em Nature ↗
Technology