Smooth Multi-Robot Routing in Real Warehouses

Most companies get multi-robot navigation wrong in one specific way: they plan routes like the world is a chessboard.

If you’re running (or buying) warehouse automation, AMRs, or any kind of fleet robotics, you already know the pain point. One robot is easy. Ten robots are manageable. A hundred robots? That’s where “they’ll just avoid each other” turns into traffic jams, near-misses, and throughput that never matches the spreadsheet.

This post is part of our AI in Robotics & Automation series, and it focuses on a research direction that’s especially relevant right now: multi-agent path finding in continuous environments—planning smooth, physically drivable paths for many robots at once, while preventing collisions in both space and time. A recent approach called Continuous-Environment Conflict-Based Search (CE-CBS) shows a practical way to get there by combining sampling-based motion planning with conflict-resolution logic designed for fleets.

Why multi-agent path finding (MAPF) is still the bottleneck

Answer first: Fleet performance is limited less by robot speed and more by coordination—because every avoidable stop, deadlock, and detour compounds across the whole system.

In a typical warehouse, robots share narrow aisles, merge at intersections, queue at pick/pack stations, and pass humans or other moving equipment. Your fleet might “work” in demos but degrade under real load because:

• Local avoidance doesn’t optimize globally. Two robots may politely yield to each other… and create a third robot’s blockage.
• Discrete planning hides real physics. Grid-based routes produce sharp turns and corner cutting that are hard to execute cleanly.
• Time is the missing dimension. “Don’t be in the same place” isn’t enough; it must be “don’t be in the same place at the same time.”

That’s the core of multi-agent path finding (MAPF): multiple agents (robots, vehicles, even people) each need to reach their goals simultaneously, without collisions, ideally with short travel time and high throughput.

Warehouse buyers usually feel MAPF as business metrics: missed cycle times, reduced pick rates, congestion near high-demand SKUs, and the need to cap fleet size earlier than expected.

Discrete grids vs. continuous reality: where plans break

Answer first: Grid-based MAPF is great for theory and simulations, but continuous-space planning is what makes routes drivable and stable at speed.

Classic MAPF research often models the environment as a graph (commonly a 2D grid). Vertices represent positions; edges represent moves. That abstraction makes planning tractable and lets algorithms prove optimality.

But real facilities aren’t tidy grids:

• A robot’s footprint isn’t a point; it’s a body with radius/shape.
• Turns have curvature constraints; sharp 90° turns can force slowdowns.
• Obstacles aren’t perfectly aligned to a grid.
• Navigation stacks must respect acceleration limits and controller behavior.

Here’s the practical issue I see repeatedly: a planner can produce a “shortest path” on a grid that is not the fastest path for a physical robot—because it demands aggressive turning or frequent stop-and-turn behavior.

A continuous-space approach aims for something more operationally honest: paths that look like what robots actually drive—smooth curves around obstacles, with timing that avoids collisions.

The building blocks: sampling-based planning plus smoothing

Answer first: Sampling-based algorithms approximate continuous environments well, and spline smoothing turns those approximations into paths robots can follow without constant braking.

To plan in a continuous environment, you can’t enumerate “all possible positions.” Instead, you sample points in space and connect them to form a navigable structure.

Why sampling works (and why it’s used in industry)

Sampling-based planners such as RRT* randomly sample coordinates in the free space and progressively build a tree of feasible connections. Over time, you get paths that approach near-optimal length under certain assumptions.

For robotics teams, the appeal is straightforward:

• Handles complex obstacle shapes
• Scales better than exhaustive continuous search
• Adapts to varying map resolutions

Why smoothing matters more than people think

A raw RRT* path is typically a polyline—many short segments with sharp direction changes. Real robots can follow it, but they’ll often:

• slow down at each “corner,”
• oscillate during tracking,
• or require overly conservative speed limits.

The CE-CBS approach smooths each agent’s path using B-splines (piecewise polynomial curves). This produces curvature-friendly trajectories—less stop-and-go, more stable control, and usually better throughput.

A clean rule of thumb: if your plan is not smooth, your robots will create their own smoothness by slowing down. That’s throughput you’re donating to physics.

CE-CBS: resolving conflicts in continuous space and time

Answer first: CE-CBS plans smooth paths per robot, then uses conflict-based search logic to detect collisions over time and add targeted constraints until the fleet is conflict-free.

The heart of the research is an extension of Conflict-Based Search (CBS) into continuous environments, called Continuous-Environment CBS (CE-CBS).

CBS is popular in discrete MAPF because it’s conceptually clean: plan individually, detect conflicts, then resolve conflicts by branching on constraints. CE-CBS keeps that spirit but adapts it to continuous geometry and timing.

Level 1: plan each robot’s smooth path

Each robot:

1. Finds a path in continuous space using RRT*
2. Smooths the path using B-splines

This creates a candidate solution for each agent that is feasible relative to static obstacles.

Level 2: detect conflicts and add constraints

CE-CBS then checks whether any two robots overlap at any time, modeling robots as rigid circular bodies (a reasonable first-order approximation for many AMRs).

When a conflict is found, the algorithm:

• Creates a constraint that forbids one robot from occupying a particular region during a specific time interval
• Branches the search tree by trying constraints on either robot involved in the conflict
• Re-plans only the affected robot(s), then rechecks the whole set

This repeats until all conflicts are removed.

The important practical detail: constraints are surgical—they target where and when the conflict occurred. That’s different from naive strategies like globally slowing robots down or inflating safety buffers everywhere.

What this buys you operationally

If you’re thinking in warehouse KPIs, CE-CBS-style planning aims to reduce:

• Intersection deadlocks
• “accordion braking” in dense robot traffic
• detours caused by conservative local avoidance
• variability in travel time (which hurts station scheduling)

It won’t eliminate every situation—narrow corridors may still require yielding or speed reductions—but it raises the ceiling on what coordinated fleets can do.

Where this fits in warehouse and manufacturing automation

Answer first: Continuous-space MAPF is most valuable where you need predictable, high-throughput robot flow—dense fleets, narrow geometry, and strict service-level timing.

You’ll see the biggest impact in environments like:

High-density AMR fleets in fulfillment

Peak season planning is a real December problem: order volume spikes, labor is tight, and facilities push automation harder than usual. In those conditions, congestion isn’t a “rare edge case”—it’s the norm.

Smooth, coordinated routing helps when:

• many robots share a few chokepoints (induct, pack-out, charging)
• pick density causes repeated crossing flows
• you need consistent cycle times to hit shipping cutoffs

Manufacturing intralogistics (line-side delivery)

In manufacturing, variability is expensive. If a robot delivery arrives late, a line can starve. Continuous-time conflict reasoning matters because:

• the timing of aisle occupancy can be scheduled like a resource
• predictable motion reduces the need for large buffers in production planning

Mixed environments with humans

Most sites are not “lights out.” Smooth paths reduce abrupt turns and stop-start behavior, which can make robot motion more legible to nearby workers. That doesn’t replace safety standards, but it helps reduce awkward interactions that kill productivity.

Practical adoption: what to ask vendors (or your robotics team)

Answer first: If you want the benefits of multi-agent path finding, you need to evaluate curvature, timing, and conflict resolution—not just collision avoidance.

When you’re assessing an AMR fleet manager or robotics platform, these are the questions that cut through marketing:

1. Do routes include curvature constraints or smoothing?

   • Ask to see paths at operating speed, not just in a map viewer.

2. Is collision avoidance reactive-only, or does it include time-based coordination?

   • Reactive approaches can be safe but still produce gridlock.

3. How does the system handle chokepoints?

   • Look for reservation systems, time windows, or explicit intersection management.

4. What’s the policy for deadlocks and livelocks?

   • “Robots will negotiate” isn’t a policy.

5. Can the planner re-optimize locally without re-planning the entire fleet?

   • CE-CBS-style constraints suggest a structured way to update plans.

A simple pilot metric that reveals coordination quality

If you’re piloting a fleet, track time spent below nominal speed due to interactions (not obstacles). Even without fancy instrumentation, you can approximate it from robot logs and commanded velocities.

If that number climbs sharply as you add robots, you’re looking at a coordination limit—not a hardware problem.

People also ask: does continuous MAPF replace local avoidance?

Answer first: No—continuous MAPF and local avoidance should work together, with MAPF handling global coordination and local avoidance covering uncertainty.

Real deployments need layers:

• Global planning / scheduling: assigns tasks and generates coordinated routes and timings
• Multi-agent path finding: prevents predictable conflicts and congestion
• Local obstacle avoidance: reacts to humans, dropped boxes, and unexpected blockages

I’m opinionated here: if you rely on local avoidance to solve what is fundamentally a scheduling problem, you’ll pay for it with throughput.

What to do next if you’re scaling a robot fleet

Smooth, continuous multi-agent path finding is becoming a foundational capability in AI-powered robotics and automation, especially as facilities push toward higher robot density and tighter delivery promises.

If you’re planning a new deployment or expanding an existing one, the next step is to audit your navigation stack for two things: path quality (smoothness and speed feasibility) and time-based conflict handling. Those are the levers that separate “safe demos” from stable, high-throughput operations.

If your robots are still driving like they’re on a grid, you’re leaving capacity on the floor. What would your throughput look like if intersections and chokepoints were scheduled as carefully as labor and inventory—rather than improvised in real time?