Hierarchical Trajectory Proposal Network for Fast Diverse Path Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous navigation systems face inefficiencies in predicting diverse trajectories due to increased computation time and efforts, especially in dynamic environments like urban areas with temporary changes, and existing models often rely on simplifying assumptions that do not account for real-time environmental properties.
Innovation Solution
A hierarchical 2-stage Neural Network architecture is employed, comprising a Trajectory Proposal Network (TPNet) and a Trajectory Sampling Network (TSNet), which uses sensor data and occupancy maps to predict diverse traversable regions and fine-grained trajectories, leveraging multiple choice learning and RRT techniques for efficient trajectory planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional trajectory planning techniques based on hypothesizing/sampling multiple goal locations followed by computation of trajectories are used, then diverse trajectory prediction capability is achieved, but computation time and efforts increase significantly
Solution Approach 1:
The trajectory planning problem is segmented into two independent stages: (1) traversable region proposal stage that identifies feasible regions without computing complete trajectories, and (2) trajectory sampling stage that generates diverse trajectories only within these pre-identified regions. This segmentation avoids the computational burden of computing trajectories to all possible goal locations while maintaining diverse trajectory prediction capability.
Solution Approach 2:
The system performs preliminary action by first proposing traversable regions before generating trajectories. The Trajectory Proposal Network (TPNet) pre-identifies feasible regions in the environment, and then the Trajectory Sampler Network (TSNet) efficiently samples diverse trajectories within these regions. This preliminary region proposal action significantly reduces the search space for subsequent trajectory generation, thereby reducing computation time.
2Measurement precision
If existing trajectory planning techniques based on neural networks predict perceptual input to continuous space control actions, then path prediction capability is achieved, but efficiency in predicting multiple trajectory proposals is poor
Solution Approach 1:
The system transitions from predicting trajectories in continuous space to predicting trajectories in a discretized grid-based traversable region space. The TPNet outputs a grid map where each cell represents a traversable region, and the TSNet samples trajectories within these discrete regions. This dimensional change from continuous to discrete space enables efficient parallel processing and multiple trajectory proposals while maintaining path prediction accuracy.
Solution Approach 2:
The system extracts the traversable region information from the complex sensor input and separates it from the trajectory generation process. The TPNet extracts and outputs only the essential traversable region proposals, which are then used by the TSNet to generate multiple trajectory proposals. This extraction approach simplifies the prediction task and improves efficiency by focusing on key spatial information rather than computing all trajectory details in one step.
3Device complexity
If simplifying assumptions on the predictive model of Ego-Vehicle motion or prior knowledge such as road topologies are assumed to be available, then model complexity is reduced, but accuracy in dynamic environments with temporary changes deteriorates
Solution Approach 1:
The system employs self-service by automatically learning environmental features and road topologies from sensor data without requiring pre-programmed prior knowledge. The neural networks learn the structure of traversable regions and valid trajectory patterns directly from training data, enabling the system to adapt to different environments including dynamic urban areas with temporary changes, while maintaining reasonable model complexity.
Data Source
Figure 1
Figure 2
Figure 3a
AI summary
The disclosure herein generally relates to the field of autonomous navigation, and, more particularly, to a hierarchical network based diverse trajectory proposal for autonomous navigation. The hierarchical 2-stage neural network architecture maps the perceived surroundings to diverse trajectories in the form of trajectory waypoints, that an autonomous navigation system can choose to navigate/traverse. The first stage of the disclosed hierarchical 2-stage Neural Network architecture is a Trajectory Proposal Network (208) which generates a set of diverse traversable regions in an environment which can be occupied by the autonomous navigation system in the future. The second stage is a Trajectory Sampling network (214) which predicts a fine-grained trajectory/trajectory waypoint over the diverse traversable regions proposed by Trajectory Proposal Network.