In the field of high-end equipment manufacturing, safe, reliable, and rapidly adaptive navigation capabilities are crucial for robots to be successfully applied in a wider range of real-world environments. However, over the past few decades, although robotics experts and computer scientists have introduced various computational techniques for robot navigation, many of these methods perform poorly in dynamic, cluttered, or narrow-path environments.
Recently, researchers from the Zhejiang University Huzhou Research Institute in China have made significant progress by proposing a new robot navigation method that combines deep neural networks with classical optimization techniques. The related scheme was published in Science Robotics, aiming to artificially replicate human pathfinding capabilities.
The first author of the paper, Han Zhichao, told Tech Xplore that the research motivation was to develop a trajectory planner capable of robust operation in arbitrarily complex environments while respecting the robot's non-holonomic constraints. Drawing inspiration from human reasoning, humans can intuitively identify rough feasible paths in complex environments, even if those paths are not optimal or fully safe. To this end, the team implemented a lightweight neural network to simulate this process.
However, while artificial neural networks perform well in many tasks, their predictions are often difficult to interpret, and many techniques based on these networks struggle to generalize across a wide range of scenarios. To overcome these limitations, Han Zhichao and his colleagues combined deep neural networks with a newly developed spatio-temporal trajectory optimizer to further refine the trajectories and paths generated by the neural network.
Professor Han explained that the proposed hierarchical planning framework has two key objectives. First, in the initial path planning stage, a learning-based method is used to reproduce humans' ability to "immediately" grasp feasible routes in the environment, ensuring stable and predictable planning time. Second, the framework ensures that the initial paths generated by the neural network can be converted into smooth motion commands executable by real robots, relying on specialized numerical optimization techniques designed to improve trajectories and paths.
"The core idea of the algorithm is to imitate the human planning process, where past experience plays a vital role in path planning," Professor Han explained. The algorithm learns from a large dataset of expert demonstrations, distilling prior knowledge into the network. Moreover, the neural planner operates directly in the same image domain as the environment representation, accelerating training speed and enhancing convergence performance.
Preliminary tests show that the pathfinding method developed by Han Zhichao and his colleagues is more stable than previous neural network-based approaches. Regardless of environmental complexity, it can reliably output paths for the robot within a fixed and predictable time frame. In contrast, many traditional planning methods require extensive online search, causing delays in pathfinding in dynamic or challenging environments and slowing robot navigation.
Professor Han stated that this research effectively combines classical numerical optimization with deep neural networks, leveraging the strengths of both while compensating for their weaknesses. Deep networks are efficient but lack completeness guarantees, while classical methods are complete but performance depends heavily on initialization. After integration, the system achieves stable and high-quality spatio-temporal trajectory generation in challenging environments.
Currently, the pathfinding method proposed by the research team is scheduled for experimental validation on a wider variety of robot platforms in the near future. In the long term, it holds promise for enhancing robots' execution capabilities in complex tasks such as search and rescue operations, logistics missions, and exploration in dynamic environments. Professor Han also mentioned that the team plans to address sim-to-real transfer challenges by improving simulation fidelity and enhancing perception robustness, ensuring that robots can operate safely, reliably, and predictably in diverse and complex real-world settings, ultimately achieving seamless integration with human daily life and industrial applications.
