Adaptive Dynamic Window Approach for Robot Navigation in Disturbance Vector Fields

Reliable autonomous robot navigation in dynamic environments with external disturbances remains challenging. The Dynamic Window Approach (DWA) generates collision-free trajectories by optimizing over sensor observations and motion constraints. However, the DWA lacks the ability to account for unpredictable disturbances, making robot navigation unreliable. We propose an enhanced planning and control approach that incorporates learned disturbance models to improve adaptability. Our key idea is to represent disturbances as parametric vector fields. By learning the vector field online, we capture environmental flows to be leveraged during planning. We integrate the learned model into the DWA objective to generate optimized trajectories through disturbance flows while avoiding obstacles. The proposed adaptive planning framework is validated in simulations and real-world experiments with ground and aquatic robots. Different case studies demonstrate the approach's ability to produce smooth, collision-free robot navigation in varied disturbance fields and environments. Compared to the standard DWA, our planner handles uncertainties and changing conditions better by learning online. Thus, this disturbance-incorporated planning enables more reliable autonomous navigation in uncertain, dynamic environments.