MarineFormer: A Spatio-Temporal Attention Model for USV Navigation in Dynamic Marine Environments (2410.13973v3)

Abstract: Navigating autonomously in marine environments including dynamic and static obstacles, and strong flow disturbances, such as in high-flow rivers, poses significant challenges for USVs. To address these challenges, we propose a novel methodology that leverages two types of attention: spatial attention, which learns to integrate diverse environmental factors and sensory information into navigation decisions, and temporal attention within a transformer framework to account for the dynamic, continuously changing nature of the environment. We devise MarineFormer, a Trans${\bf \text{former}}$-based navigation policy for dynamic ${\bf \text{Marine}}$ environments, trained end-to-end through reinforcement learning (RL). At its core, MarineFormer uses graph attention to capture spatial information and a transformer architecture to process temporal sequences in an environment that simulates a 2D turbulent marine condition involving multiple static and dynamic obstacles. We extensively evaluate the performance of the proposed method versus the state-of-the-art methods, as well as other classical planners. Our approach outperforms the state-of-the-art by nearly $20\%$ in episode completion success rate and additionally enhances the USV's path length efficiency.

• The paper introduces MarineFormer, a transformer-based RL model that integrates spatio-temporal attention for enhanced USV navigation and collision avoidance.
• It leverages multi-head cross-attention and a tailored reward function to effectively process environmental features and outperform baseline methods by 20%.
• Ablation studies underscore the value of modeling dynamic obstacles and local current flows, marking a significant advance in autonomous marine navigation.

MarineFormer: A Transformer-based Navigation Policy for Marine Environments

The paper presents "MarineFormer," a novel approach leveraging transformer-based models for collision avoidance in marine environments with Unmanned Surface Vehicles (USVs). The paper addresses the complexities of navigation in dynamic and static obstacle-rich environments influenced by strong current flows. Standard navigation protocols fall short in these conditions due to intricate interactions between the environment and the vehicle.

Methodology

MarineFormer introduces a transformer architecture within a reinforcement learning (RL) framework to comprehend spatially and temporally variable disturbances. The authors enhance the capability of USVs to maintain safe navigation by refining a temporal function specific to marine conditions. The model incorporates spatio-temporal graph attention mechanisms that enable the USV to process sequences and attend to environmental components effectively.

Specifically, MarineFormer leverages three main input features: the ego-state of the USV, static and dynamic obstacle states, and local current flow observations. The dynamic obstacle predictions incorporate future trajectories, enhancing decision-making by predicting potential collision paths. The current flow information is captured on a grid surrounding the USV, using convolution layers to prepare this data for input into the transformer model.

Network Architecture

The transformer architecture, chosen for its superior ability to handle long-term dependencies, structures the input features into an attention-enabled graph. The temporal progression through the task is managed by a multi-head cross-attention mechanism layered over the usual sequence embeddings. The network's robustness is further enhanced by the inclusion of residual connections and layer normalization.

Reward Function

To guide the navigation towards collision-free and efficient trajectories, the paper designs a comprehensive reward structure. It entails proximity rewards to avoid static and dynamic obstacles and penalizes trajectories trapped by flow-induced singularities. The function considers future trajectory predictions, enabling the USV to navigate proactively rather than reactively.

Results and Comparisons

MarineFormer achieves a notable improvement over baseline methods, with a 20% increase in success rate compared to state-of-the-art models. The robust architecture allows the model to remain unaffected by vortices and sinks, showcasing superior adaptability. Ablation studies demonstrate the critical role of various components such as similarity scoring among dynamic obstacles and inclusion of current flow data on the policy's success.

Implications and Future Work

The implications of this research extend to enhancing autonomous navigation systems for marine applications, presenting a model adept at handling complex interactions in uncertain, dynamic environments. Future developments may include incorporating real-world dynamic obstacle behaviors and refining flow disturbance modeling, ensuring broader applicability and increased robustness against sensor noise.

In summary, MarineFormer establishes a significant contribution to autonomous marine navigation, using advanced deep learning methods to effectively solve critical real-world challenges associated with USV operations.