Robots can find their way around a warehouse — but what happens when two of them want the same shelf at the same time?
image from grok
CREST addresses the coordination bottleneck in warehouse robotics rather than navigation—where multiple robots competing for the same shelf creates idle time and unnecessary shelf switches. Instead of the traditional MAPF-DECOMP approach that commits to trajectories upfront, CREST releases trajectory constraints during execution, allowing continuous adaptive decision-making. In simulation across diverse warehouse layouts, CREST achieved up to 44.4% reduction in shelf switches, 40.5% reduction in agent travel, and 33.3% reduction in makespan, though these results remain unvalidated against real warehouse operations.
The problem CREST addresses is not navigation. Warehouse robots can already find their way around a facility. The bottleneck is what happens when a robot needs a shelf that another robot is sitting on top of. In a large automated warehouse, hundreds of robots move constantly, and shelves get relocated constantly. When one robot needs a shelf that another robot is using, the coordination problem is significant.
The existing approach, MAPF-DECOMP, computes collision-free trajectories first and then assigns robots to execute them. The researchers describe this as producing "strict trajectory dependencies" that leave robots idle and trigger unnecessary shelf switches. CREST takes a different approach: instead of committing to a trajectory upfront, it releases trajectory constraints during execution, allowing robots to make continuous decisions as conditions change. It uses three complementary constraint-release strategies: Single Trajectory Replanning, Dependency Switching, and Group Trajectory Replanning. These are designed to reduce agent idle time and unnecessary shelf switching. The code is on GitHub.
In simulation across diverse warehouse layouts, CREST reduced shelf switches by up to 44.4 percent, agent travel by up to 40.5 percent, and makespan by up to 33.3 percent, with greater benefits under high lift and place overhead. These numbers come with the usual benchmark caveat: they reflect controlled experiments, not a live distribution floor.
The authors are honest about the gap themselves. The paper notes that CREST was evaluated in a controlled simulation environment, not a real warehouse. A real distribution floor at 3 a.m. involves obstacles, sensor noise, and edge cases that a benchmark does not fully capture. No commercial warehouse operator is named as a partner or evaluator. The paper was accepted March 8, 2026 for IEEE RAL.
Jiaqi Tan, the corresponding author, is a researcher at Simon Fraser University working in the multi-agent pathfinding space. The co-author team spans SFU, HEC Montreal, and Purdue. Their work builds on prior DD-MAPD research by Li and Ma, also published in IEEE RAL. This is a well-defined academic problem with real industrial relevance — not a prototype demo.
The commercial case for this kind of coordination was built by Kiva Systems, now Amazon Robotics, more than a decade ago. Amazon has been running multi-robot coordination at scale in its fulfillment centers since the early 2010s. The company does not publish operational benchmarks for its internal systems. Any new algorithmic result has to argue against that operational experience, not just a simulation baseline. The 44-percent shelf-switch reduction is a genuine improvement over the prior algorithm. It is not clear it is a 44-percent improvement over what Amazon is already doing.
This is the familiar pattern in warehouse robotics research: the benchmark numbers are real, and the deployment gap is also real. CREST is a cleaner execution layer for a well-defined coordination problem. That is worth publishing. It is not yet a deployment story.
Story entered the newsroom
Research completed — 0 sources registered. CREST is a real, peer-reviewed algorithm (IEEE RAL, accepted March 8 2026) from Jiaqi Tan et al. at Simon Fraser University. It improves multi-robot w
Draft (452 words)
Approved for publication
Headline selected: The Warehouse Problem Isn't Navigation — It's Coordination
Published (472 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 2d ago · 4 min read
Robotics · 2d ago · 4 min read
