LDP: A Local Diffusion Planner for Efficient Robot Navigation and Collision Avoidance

TL;DR

The paper addresses the challenge of local robot navigation under dynamic obstacles by modeling the policy as a multimodal conditional distribution and learning from diverse expert data. It introduces LDP, a diffusion-based local planner that conditions on costmaps, goals, and global paths, trained with a diffusion-denoising objective and classifier-free guidance to generate action sequences given observations. The authors release expert-policy data across three scenarios with two preferences and demonstrate that LDP achieves superior navigation performance, robustness, and zero-shot generalization in simulation and real-world Ackermann robots, surpassing baselines like LSTM-GMM, IBC, and DT. The work advances practical collision avoidance by combining diffusion modeling with global-path guidance and mixed-preference data, with future directions including richer data and faster sampling methods to enable real-time deployment.

Abstract

The conditional diffusion model has been demonstrated as an efficient tool for learning robot policies, owing to its advancement to accurately model the conditional distribution of policies. The intricate nature of real-world scenarios, characterized by dynamic obstacles and maze-like structures, underscores the complexity of robot local navigation decision-making as a conditional distribution problem. Nevertheless, leveraging the diffusion model for robot local navigation is not trivial and encounters several under-explored challenges: (1) Data Urgency. The complex conditional distribution in local navigation needs training data to include diverse policy in diverse real-world scenarios; (2) Myopic Observation. Due to the diversity of the perception scenarios, diffusion decisions based on the local perspective of robots may prove suboptimal for completing the entire task, as they often lack foresight. In certain scenarios requiring detours, the robot may become trapped. To address these issues, our approach begins with an exploration of a diverse data generation mechanism that encompasses multiple agents exhibiting distinct preferences through target selection informed by integrated global-local insights. Then, based on this diverse training data, a diffusion agent is obtained, capable of excellent collision avoidance in diverse scenarios. Subsequently, we augment our Local Diffusion Planner, also known as LDP by incorporating global observations in a lightweight manner. This enhancement broadens the observational scope of LDP, effectively mitigating the risk of becoming ensnared in local optima and promoting more robust navigational decisions.

Figures (6)

• Figure 1: The diagram illustrates the execution of our method. Obstacles are denoted by black circles and rectangles, while the trajectories of pedestrians are represented by green circles. The navigation target is marked by a yellow pentagram, and a brown dashed line delineates the global path from the robot's starting point to its target. This system utilizes laser data, target point information (goal), and the global path to generate local action sequences in a classifier-free guidance approach.
• Figure 2: An in-depth depiction of the entire process and the architecture of the local diffusion planner. The circles, marked with different colors and denoted as $\tau$, represent expert data gathered from diverse scenarios, each reflecting various preferences through distinct levels of transparency. At time step $t$, the planner takes in observations $\Vec{O}_t$ from the past $T_o$ steps and predicts the action sequence $\Vec{A}_t$ for the next $T_a$ steps. During the training process, the loss is computed based on DDPM noise prediction over the entire sequence $T$. In the inference process, an action sequence can be generated for every $K$ iterations of denoising.
• Figure 3: Four different simulation scenarios are displayed. The black rectangles and circles are obstacles, the green dots represent pedestrian trajectories, and the blue box on the right shows the robot's local sensor map.
• Figure 4: Global Path Influence: Navigation Success vs. Failure in One Scene
• Figure 5: Performance of policies learned under expert data with different preferences.