Long-Horizon Geometry-Aware Navigation among Polytopes via MILP-MPC and Minkowski-Based CBFs
A new hierarchical AI planning system solves the 'local minimum' trap, letting robots navigate tight, complex spaces safely.
A research team from institutions including Boston University and Texas A&M has published a novel framework for autonomous robot navigation in cluttered, non-convex environments. The core innovation is a two-layer system that marries long-term strategic planning with instantaneous, geometry-aware safety. At the high level, a Mixed-Integer Linear Programming Model Predictive Control (MILP-MPC) planner charts a nominal course through polytopic (angular) obstacles, treating the robot as a point for computational speed. This solves the critical problem of robots getting stuck in 'local minima,' like dead ends in a maze.
At the low level, a safety filter using a Minkowski-difference-based Control Barrier Function (CBF) takes over. This component is key: it uses the exact signed distance between the robot's true polygonal shape and obstacles to enforce collision avoidance in real-time, respecting the robot's actual geometry and dynamics. The team extended this to a High-Order CBF (HOCBF) to handle more complex motion models. In tests with single- and double-integrator dynamics in U-shaped and maze settings, the framework proved capable of safe, real-time navigation where purely reactive systems would fail.
- Uses a hierarchical MILP-MPC planner for long-range strategy to avoid navigation dead-ends (local minima).
- Employs a Minkowski-based Control Barrier Function (CBF) as a safety filter for real-time, exact-shape collision avoidance.
- Demonstrated successful real-time navigation in complex U-shaped and maze-like environments with polytopic obstacles.
Why It Matters
This enables more reliable autonomous robots for logistics in tight warehouses, search-and-rescue in rubble, and other complex real-world navigation tasks.