.png){kind=icon}
.png){kind=icon}
Return
End-to-End Autonomous Navigation: A Deep Unfolding and Reinforcement Learning-Based Approach
.png){kind=icon}
PRE
AI
X
G
H
DOI：10.1109/tie.2026.3679837.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Autonomous navigation demands the ability to operate in unknown and unstructured environments using only onboard, limited-range sensor data, without reliance on prior maps or global information. To address this challenge, this article introduces a novel end-to-end navigation framework composed of perception and planning modules. By adopting a modular end-to-end paradigm, the framework avoids the compounding errors inherent in traditional, decoupled pipelines. The perception component, Perception Net, is an interpretable deep unfolding network that efficiently processes raw point clouds into a compact feature representation of arbitrary obstacles, exhibiting significant advantages in both efficiency and accuracy. The planning component is a reinforcement learning planner featuring a velocity-adaptive Bézier action space, which learns a smooth and kinematically feasible navigation policy endowed with a forward-looking capability to handle nonconvex environments. The agent’s decisions are based solely on its own state, goal coordinates, and real-time perception features. Extensive simulations demonstrate the framework’s high success rate and robust generalization to complex scenarios. The effectiveness of this end-to-end framework is further validated through real-world field tests on an autonomous surface vehicle platform, demonstrating its capability to operate as a map-less and forward-looking system for autonomous navigation in complex environments.
Keywords:
Deep unfolding
end-to-end autonomous navigation
motion planning
reinforcement learning (RL)
Journal
IF:
7.2
Papers:
1.8W
Citations:
9.8W
Organization
Cited Papers