LWAT: Learning Width-Aware Trajectory-Level Actions for Real-Time Robot Motion Planning
Abstract
Generating smooth and safe trajectories from onboard sensor data remains a fundamental challenge for mobile robots operating in cluttered and dynamic environments. Conventional deep reinforcement learning planners commonly output instantaneous control actions, which often leads to short-sighted decisions, fragmented motions, and limited geometric reasoning over future trajectory evolution. To address these limitations, we propose a Learning Width-Aware Trajectory-Level Actions (LWAT) framework that formulates motion planning as a trajectory-level prediction problem rather than a step-wise control problem. A lightweight policy predicts structured multi-node trajectory actions, which are transformed into continuous, width-aware, and executable trajectories through a differentiable trajectory parameterization network (TPN). This formulation enables the policy to reason over complete local trajectories from the current robot state to the local goal within the current LiDAR observation while explicitly accounting for robot footprint and safety margins. Extensive simulations and real-world experiments demonstrate that LWAT reliably handles cluttered dynamic environments and narrow-passage navigation, achieving a 95% success rate and improving average clearance by 14.3% over baseline, while running in real time without GPU acceleration.