ML-Assisted Warehouse Robot Navigation Using UCS

ML predicts real-time path costs — traffic, obstacles, battery — and Uniform Cost Search finds the smartest route, not just the shortest one. Adaptive navigation for autonomous warehouse robots.

Comment

Inspiration

Most warehouse robots are still routed using the shortest physical path — but in a real warehouse, the shortest path isn't always the smartest one. Traffic congestion, obstacle density, and battery level constantly change, and a route that looks optimal on paper can become inefficient seconds after the robot starts moving. We wanted to build a navigation system that predicts how costly a route will actually be before committing to it, instead of just measuring distance.

What it does

Our system combines a Machine Learning cost-prediction model with Uniform Cost Search (UCS) path planning. Instead of relying on fixed edge costs, a Random Forest model predicts the traversal cost of each path segment using four live features — distance, obstacle density, traffic, and battery level. These predicted costs are fed into UCS, which computes the actual minimum-cost route rather than the shortest-distance one. The system is designed to re-route adaptively as warehouse conditions change.

How we built it

  • ML Cost Model: Random Forest Regression trained on 600 synthetic samples (R² = 0.95, MAE = 1.44, RMSE = 1.74)
  • Path Planning: Uniform Cost Search implemented and validated on a toy graph, with adaptive re-routing logic
  • Simulation Environment: A Webots warehouse environment with racks, start/goal markers, and dynamic obstacles, running an autonomous robot with reactive obstacle avoidance
  • Pilot Validation: 25 automated Webots trials to establish a baseline success rate for the current build
  • IP: We've drafted a full patent specification for the extended architecture this system is designed to scale into — an Adaptive Navigation Intelligence (ANI) Engine incorporating an Operational Experience Repository, Navigation Confidence Index, and Predictive Congestion Estimator. It's currently under institutional review.

Challenges we ran into

  • Our ML model's feature importance showed battery level contributing very little (0.03) compared to distance, obstacle density, and traffic — weaker than we expected, and something we're actively working to strengthen in the training data
  • Fusing the ML-predicted costs directly into live UCS-driven robot navigation in Webots (rather than running each component independently) turned out to be more involved than validating them separately
  • Our 25-trial pilot run, using the current reactive-navigation build, returned a 44% success rate — a useful baseline, but a clear signal of where the system still needs work before the full ML+UCS pipeline is fused end-to-end

Accomplishments that we're proud of

  • Built and validated three independent components of the pipeline: the ML cost model, the UCS algorithm, and the Webots simulation environment
  • Achieved a high-accuracy cost prediction model (R² = 0.95) using an explainable, fast-to-train approach instead of a slower reinforcement-learning method
  • Drafted a complete patent specification for the broader system architecture

What we learned

  • Supervised regression + classical search can match the adaptivity goals of reinforcement learning approaches while staying faster to train and easier to explain
  • Validating components independently is a useful first step, but true end-to-end system behavior only shows up once everything is fused and tested together
  • Being transparent about a weak pilot result (44%) is more valuable for real engineering progress than hiding it

What's next

  • Fuse ML-predicted costs into live UCS-driven waypoint navigation in Webots, replacing the current reactive-only mode
  • Strengthen the battery-cost signal in the ML model
  • Deploy on an ESP32-based hardware prototype
  • File the patent specification once institutional review clears
  • Implement the extended ANI Engine modules: Operational Experience Repository, Navigation Confidence Index, and Predictive Congestion Estimator
  • Scale from single-robot validation to multi-robot fleet coordination

Built With

Share this project:

Updates

Submission history