Abstract
Robot navigation in complex environments requires balancing long-term goal reaching with immediate obstacle avoidance. Traditional local planners often suffer from "myopic observation," leading to local optima in maze-like structures or failure in multimodal scenarios. We present the Local Diffusion Planner (LDP), which leverages conditional Denoising Diffusion Probabilistic Models (DDPM) to generate safe, goal-oriented action sequences. By conditioning the diffusion process on local sensor data and a global path using Feature-wise Linear Modulation (FiLM), LDP effectively navigates intricate environments. Our results demonstrate that LDP significantly outperforms state-of-the-art baselines in success rate and generalization across diverse simulated and real-world scenarios.
Introduction
Autonomous robot navigation is typically decomposed into global planning for long-range pathfinding and local planning for reactive obstacle avoidance. However, local planners frequently encounter the "myopic observation" problem: they lack the foresight to avoid dead-ends or navigate complex geometries that aren't immediately visible. Furthermore, real-world navigation is inherently multimodal; for instance, a robot might pass an obstacle on either the left or right. Standard regression-based imitation learning often fails here by "averaging" these modes, potentially leading to collisions.
LDP addresses these issues by treating local planning as a conditional sequence generation task. By incorporating the global path as a guiding condition, the local planner gains the necessary "foresight" to align immediate actions with long-term objectives.
Methodology
LDP utilizes Denoising Diffusion Probabilistic Models (DDPM) to iteratively refine a sequence of future actions from Gaussian noise. This generative approach allows the model to capture the multimodal nature of navigation trajectories.
Conditional Diffusion and FiLM
The core of LDP is a conditional diffusion process. Unlike traditional models that predict a single next action, LDP generates a full trajectory (action sequence) $\vec{A}_t$. This generation is conditioned on the robot's current state, including laser data, the local goal, and the global path.
To integrate these heterogeneous inputs, LDP employs Feature-wise Linear Modulation (FiLM). FiLM allows the model to dynamically scale and shift intermediate features based on the conditioning information, ensuring that the generated path is both feasible and goal-oriented.
Classifier-Free Guidance
During inference, the model uses a guidance scale ($\omega$) to balance between following the learned distribution and strictly adhering to current observations. This ensures the generated path is both feasible and goal-oriented, maintaining responsiveness to immediate obstacles while following the broader intent of the global path.
Experimental Setup
To train and evaluate LDP, a comprehensive expert dataset was developed across three challenging scenarios:
- Dense Static: Environments packed with stationary obstacles.
- Dynamic Pedestrian: Scenarios involving moving agents.
- Maze-like: Complex structures requiring long-term planning consistency.
Expert data was collected from a mix of "greedy" (Soft Actor-Critic) and "foresighted" (SAC guided by A*) agents. This mixed-expert approach provides a rich distribution, allowing LDP to learn both fast reactive behaviors and cautious, path-following strategies.
Results
LDP was benchmarked against several competitive baselines, including LSTM-GMM, Implicit Behavior Cloning (IBC), and Decision Transformers (DT).
- Performance in Complex Maps: In maze-like environments, LDP achieved a success rate of 91.5%, significantly higher than IBC (71.5%) and DT (68.6%). The inclusion of global path conditioning was critical in preventing the robot from getting stuck in local optima.
- Generalization: LDP demonstrated robust zero-shot generalization. When tested on an unseen "zigzag" map, it maintained a 72% success rate, whereas the best-performing baseline dropped to 53%.
- Efficiency: LDP consistently showed lower "Stuck Rates" and faster "Average Time" to reach the goal compared to baselines.
- Real-World Deployment: The system was successfully validated on an Agilex hunter2.0 outdoor robot equipped with 3D LiDAR, proving it can handle the transition from simulation to noisy, real-world hardware.
Conclusion and Future Work
LDP represents a significant step forward in combining generative modeling with robot navigation. By conditioning on global paths, it overcomes the limitations of myopic local planners. However, the iterative nature of diffusion models poses a challenge for real-time inference. Future work will focus on optimizing inference speed through flow-based or consistency models and scaling the dataset to include more diverse unstructured environments.
Summary updated