Dynamic Obstacle Avoidance through Uncertainty-Based Adaptive Planning with Diffusion

TL;DR

This work tackles collision avoidance in dynamic environments by integrating diffusion-model-based offline RL with uncertainty-aware adaptive planning. A deep ensemble inverse dynamics model provides predictive uncertainty to trigger replanning only when necessary, balancing long-horizon planning with safety and computational efficiency. Empirical results in highway-env show longer, safer trajectories and significant reductions in recomputation, controlled by a tunable threshold $\\epsilon$. The approach offers a practical, tunable framework for real-time autonomous navigation under moving obstacles, with broad applicability to robotics and autonomous vehicles.

Abstract

By framing reinforcement learning as a sequence modeling problem, recent work has enabled the use of generative models, such as diffusion models, for planning. While these models are effective in predicting long-horizon state trajectories in deterministic environments, they face challenges in dynamic settings with moving obstacles. Effective collision avoidance demands continuous monitoring and adaptive decision-making. While replanning at every timestep could ensure safety, it introduces substantial computational overhead due to the repetitive prediction of overlapping state sequences -- a process that is particularly costly with diffusion models, known for their intensive iterative sampling procedure. We propose an adaptive generative planning approach that dynamically adjusts replanning frequency based on the uncertainty of action predictions. Our method minimizes the need for frequent, computationally expensive, and redundant replanning while maintaining robust collision avoidance performance. In experiments, we obtain a 13.5% increase in the mean trajectory length and a 12.7% increase in mean reward over long-horizon planning, indicating a reduction in collision rates and an improved ability to navigate the environment safely.

Figures (3)

• Figure 1: Diffusion model generates an initial long-horizon trajectory based on the past history of states. However, in a dynamically changing environment, moving obstacles sharply increase the risk of collision. Our model proposes uncertainty-based adaptive planning to detect the risk of an impending collision and trigger an appropriate re-planning of the trajectory.
• Figure 2: Mean trajectory length (out of a maximum of 100 steps) is shown for three approaches: (a) Adaptive replanning based on uncertainty estimates (ours) with $\epsilon=0.1$, (b) Decision Diffuser (DD) long-horizon, i.e. with no replanning and (c) DD with continuous replanning at every time step. Results are shown for $n=10$ episodes with random initialization.
• Figure 3: This plot illustrates the impact of varying uncertainty threshold value $\epsilon$ on key performance metrics: mean trajectory length, mean reward, and collision rate. As the threshold value increases, a decline in both mean trajectory length and reward is observed, accompanied by a corresponding rise in the collision rate, highlighting the impact of lower rate of adaptive re-planning when the threshold is higher.