Neural Network Trajectory Planning for Non-Holonomic Robots
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing motion planning algorithms for non-holonomic robots in cluttered environments are computationally burdensome and prone to local minima, requiring extensive collision checking and iterative propagation of cost values.
Innovation Solution
The implementation of a neural network-based motion planning system that uses an adaptive sampling method to capture high curvature regions in a cost-to-go value map, allowing for efficient generation of collision-free trajectories for non-holonomic robots.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional lattice-based planners or RRT planners are used for non-holonomic motion planning, then the robot can navigate cluttered environments, but the computational burden increases significantly due to exhaustive random samples, collision checks, and post-processing
Solution Approach 1:
The patent replaces traditional mechanical motion planning algorithms (lattice-based planners, RRT planners) with a neural network-based system. The neural network is trained to directly predict optimal trajectories and control inputs, substituting the iterative computational mechanics of traditional planners with a learned predictive model that achieves comparable reliability at lower computational cost
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network on a dataset generated from traditional motion planners before deployment. This pre-training phase captures the essential planning knowledge, allowing the network to make rapid predictions during actual operation without requiring exhaustive collision checks or random sampling in real-time
2Manufacturing precision
If conventional approaches such as navigation functions or differential dynamic programming are used, then optimal trajectories can be computed, but the systems are vulnerable to local minima and require difficult parameter tuning
Solution Approach 1:
The patent substitutes gradient-based optimization methods (navigation functions, differential dynamic programming) with a neural network that has been trained to directly output optimal trajectories. This replacement eliminates the need for real-time parameter tuning and avoids local minima issues by leveraging the network's learned decision boundaries
Solution Approach 2:
The patent transforms the continuous parameter tuning problem into a discrete classification problem during training. By using a dataset of optimal trajectories with associated control inputs, the network learns to select appropriate control parameters without requiring manual tuning during operation
3Quantity of substance
If uniform sampling is used for training data collection, then all regions are covered, but high curvature regions in the cost-to-go value map are not adequately captured, reducing training effectiveness
Solution Approach 1:
The patent applies local quality by using adaptive sampling that concentrates training data in regions of high curvature or high gradient in the cost-to-go value map. Instead of uniform sampling, the method identifies and oversamples critical regions where accurate cost-to-go estimation is most difficult, improving overall measurement precision without increasing total data quantity proportionally
Data Source
AI summary
A system and a method for generating a trajectory of a target device from a current position to a goal position within an environment is provided. The method may include: inputting physical workspace information associated with the environment in which the target device, to a first neural network to obtain a set of weights representing a cost-to-go function that defines a cost-to-go function relating to a length of a collision-free path from one position to the goal position; configuring a second neural network based on the set of weights; identifying a next position of the target device based on the current position and a motion control input of the target; and inputting the identified next position of the target device and the goal position to the second neural network to identify the trajectory to the goal position, and the motion control input corresponding to the trajectory.