A Progressive Diffusion-Based Imitation Learning Framework for Local Motion Planning
TL;DR
A diffusion-assisted progressive imitation learning framework for a point-cloud-to-velocity local motion planner that supports improved expert-behavior reproduction within the evaluated scenario.
Abstract
Learning local navigation maneuvers from expert demonstrations is attractive when the desired behavior is difficult to encode with hand-crafted rules, but behavioral cloning (BC) can accumulate errors during closed-loop deployment. This paper proposes a diffusion-assisted progressive imitation learning framework for a point-cloud-to-velocity local motion planner. A state-action temporal diffusion model provides adversarial learning signals for a deterministic planner optimized through a critic-free horizon-level Gaussian proximal policy optimization (PPO) formulation, while diffusion-purified behavioral cloning (DP-BC) and diffusion-augmented behavioral cloning (DBC) are used for continual fine-tuning. The method is evaluated on held-out expert trajectories and ten randomized closed-loop deployment trials per configuration in a CARLA scenario. The full framework yields the lowest observed mean trajectory dynamic time warping (DTW), 0.65 ± 0.10, compared with 1.77 ± 0.23 for BC and 0.84 ± 0.23 for standalone DBC. Within the staged ablation, diffusion adversarial imitation learning (DAIL) establishes the initial reduction in trajectory discrepancy, DP-BC provides only a modest numerical reduction when added to DAIL, and DBC produces the larger incremental improvement; the lowest observed mean DTW occurs in the complete pipeline. These results support improved expert-behavior reproduction within the evaluated scenario.