车讯:缤智的对手来了 日产小型SUV专利图曝光

For robots to be successfully introduced in a wider range of real-world settings, they should be able to safely and reliably navigate rapidly changing environments. While roboticists and computer scientists have introduced a wide range of computational techniques for robot navigation over the past decades, many of them were found to perform poorly in environments that are dynamic, cluttered or characterized by narrow pathways.

Researchers at Huzhou Institute, part of Zhejiang University in China, recently introduced a new approach to robot navigation that is based on a deep neural network and classical optimization techniques. Their proposed approach, outlined in a paper published in Science Robotics, is designed to artificially replicate the pathfinding capabilities of humans.

"Our motivation was straightforward: to develop a trajectory planner that can operate robustly in arbitrarily complex environments while respecting nonholonomic constraints of robots," Zhichao Han, first author of the paper, told Tech Xplore.

"We drew inspiration from human reasoning—specifically, how people can often intuitively identify a rough path through complex environments at a glance, even if the solution is not always optimal or completely safe. To emulate this, we implemented a lightweight neural network that approximates this process."

While artificial neural networks have been found to perform well in various tasks, their predictions are often difficult to interpret. Moreover, many techniques based on these networks do not generalize well across a broad range of scenarios.

To overcome these limitations, Han and his colleagues combined a deep neural network with a newly developed spatiotemporal trajectory optimizer. This ultimately allowed them to further refine the trajectories and paths generated by the neural network.

"Our hierarchical planning framework is designed to address two key goals," said Han. "Firstly, by leveraging learning-based approaches for the initial path planning stage, we aim to reproduce the human-like ability to 'instantly' grasp a feasible route through an environment. This ensures planning times are stable and predictable."

The second goal of the team's proposed framework is to ensure that the initial paths generated by neural networks are converted into smooth motion commands that can be executed by real robots. To do this, the framework relies on numerical optimization techniques specifically aimed at improving trajectories and paths.

"The core idea is to mimic the human planning process, in which past experience plays a crucial role in path planning," explained Han. "Similarly, our algorithm learns from a large dataset of expert demonstrations, distilling this prior knowledge into the network.

"One key component is that the neural planner operates directly in the same image domain as the environment representation, which greatly accelerates both training and enhances convergence performance. Intuitively, if you ask a human to draw a path on a map, that's straightforward; asking someone to provide exact coordinate points is much less intuitive."

The pathfinding approach developed by Han and his colleagues is significantly more stable over time than previously introduced neural network-based methods. In initial tests, it was found to reliably output paths for robots within a fixed and predictable timeframe, irrespective of the complexity of a given environment.

This is a significant advantage, as many conventional planning methods need to perform extensive online searches, which can delay the pathfinding process in dynamic or challenging environments, ultimately slowing down a robot's navigation.

"We effectively combined classical numerical optimization with deep neural networks, leveraging their respective strengths while mitigating their weaknesses," said Han. "Deep networks are highly efficient but lack completeness guarantees, while classical methods are complete, but their performance tends to depend on initialization. By integrating both, our system achieves stable and high-quality spatiotemporal trajectory generation in challenging environments."

The pathfinding approach introduced by this team of researchers could soon be tested in more experiments using various robotic platforms. In the future, it could be used to improve the ability of robots to tackle different complex missions, including search and rescue operations, logistics tasks and the exploration of dynamic environments.

"Moving forward, we plan to tackle the sim-to-real transfer challenge by further improving simulation fidelity and enhancing perception robustness," added Han. "Our aim is to ensure that robots can operate safely, reliably, and predictably in diverse and complex real-world environments—ultimately achieving seamless integration into human daily life and industrial applications."

? 2025 Science X Network