ST-Graph Learning for Autonomous Vehicle Motion Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional motion planning and control systems for autonomous vehicles do not accurately account for vehicle-specific differences, relying on rule-based programming methods that require extensive experimentation and parameter tuning, and fail to leverage road test data effectively.
Innovation Solution
An autonomous driving system that generates a station-time graph based on perception data from sensors, using a multi-layered tensor and a Convolutional Neural Network to learn a discretized probability distribution, determining a path trajectory that considers obstacles and traffic rules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule-based programming methods are used for motion planning, then decisions can be made based on explicit logic, but extensive experimentation and parameter tuning are required
Solution Approach 1:
The patent replaces rule-based programming methods with a machine learning model that learns motion planning decisions directly from road test data. Instead of manually tuning parameters through extensive experimentation, the system uses neural networks to automatically learn optimal decision-making patterns, substituting mechanical parameter adjustment with data-driven learning.
Solution Approach 2:
The machine learning model performs self-learning from road test data without requiring manual parameter tuning. The system automatically improves its motion planning decisions by learning from actual driving scenarios, eliminating the need for continuous human intervention and experimentation.
2Device complexity
If same motion planning and control is applied to all types of vehicles, then system complexity is reduced, but accuracy and smoothness deteriorate under some circumstances
Solution Approach 1:
The patent incorporates vehicle-specific features as input parameters to the machine learning model, allowing the system to adapt motion planning decisions to different vehicle types, sizes, and characteristics. Each vehicle's unique properties are considered locally in the decision-making process, improving accuracy without requiring completely separate planning systems for each vehicle type.
3Reliability
If rule-based programming methods are used, then decisions follow explicit logic, but large amount of road test data is not leveraged
Solution Approach 1:
The patent replaces rule-based decision logic with a machine learning model that directly processes road test data. The system substitutes explicit programming with data-driven learning, enabling the vehicle to learn complex driving patterns and edge cases from extensive real-world testing that would be difficult to encode in rules.
Solution Approach 2:
The machine learning model creates a learned representation of optimal driving behavior by copying patterns from road test data. Instead of manually programming rules, the system replicates successful driving decisions observed in real-world scenarios, capturing nuanced behaviors that are difficult to specify explicitly.
Data Source
AI summary
In one embodiment, a data processing system for an autonomous driving vehicle (ADV) includes a processor, and a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations. The operations include generating a station-time (ST) graph based on perception data obtained from one or more sensors of the ADV, the ST graph including representing a location of an obstacle at different points in time, obtaining a tensor based on the ST graph, the tensor including a plurality of layers, the plurality of layers including a first layer having data representing one or more obstacles on a path in which the ADV is moving, applying a machine-learning model to the plurality of layers of the tensor to generate a plurality of numerical values, the plurality of numerical values defining a potential path trajectory of the ADV, and determining a path trajectory of the ADV based on the plurality of numerical values.