Interpretable Motion Planning With Cost Volumes for Autonomous Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle motion planning systems lack efficiency and accuracy in generating motion plans for complex scenarios, such as traffic light handling and interactions with multiple road users, due to the division of driving tasks into subtasks like perception, prediction, and control, which limits their ability to handle uncertainty and multimodality.
Innovation Solution
A machine-learned motion planning model that integrates end-to-end learning and interpretable intermediate representations, using a backbone network to generate a cost volume and trajectory generator to select target trajectories based on sensor and map data, enabling the autonomous vehicle to handle complex scenarios by optimizing motion planning directly from raw sensor data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If motion planning is divided into subtasks (perception, prediction, control), then the system structure becomes modular and easier to implement, but the efficiency and accuracy in generating motion plans for complex scenarios deteriorates
Solution Approach 1:
The patent combines multiple subtasks (perception, prediction, and motion planning) into a single integrated deep learning model. The model takes raw sensor data as input and directly outputs motion plans, eliminating the need for separate perception and prediction modules. This merging approach improves efficiency and accuracy in complex scenarios by allowing the model to learn optimal motion plans end-to-end without the limitations of modular processing.
2Ease of manufacture
If motion planning is divided into subtasks, then the system is easier to implement, but the ability to handle uncertainty and multimodality deteriorates
Solution Approach 1:
The integrated model merges perception, prediction, and planning functions, enabling the system to naturally handle uncertainty and multimodality through end-to-end learning. The model learns to account for uncertain factors (such as unpredictable pedestrian behavior) and multimodal outcomes (multiple valid motion plans) directly from sensor data, improving reliability in complex scenarios without requiring separate uncertainty handling modules.
3Manufacturing precision
If traditional motion planning methods are used, then the system requires costly parameter tuning, but the solution achieves accurate and efficient motion plan generation by avoiding parameter tuning
Solution Approach 1:
The patent replaces traditional mechanics-based motion planning methods (which rely on hand-crafted cost functions and extensive parameter tuning) with a deep learning-based system. The model learns optimal motion planning strategies directly from data, eliminating the need for costly parameter tuning while achieving accurate and efficient motion plan generation. This substitution transforms the planning process from a computational optimization problem to a learned pattern recognition task.
Data Source
AI summary
Systems and methods for generating motion plans including target trajectories for autonomous vehicles are provided. An autonomous vehicle may include or access a machine-learned motion planning model including a backbone network configured to generate a cost volume including data indicative of a cost associated with future locations of the autonomous vehicle. The cost volume can be generated from raw sensor data as part of motion planning for the autonomous vehicle. The backbone network can generate intermediate representations associated with object detections and objection predictions. The motion planning model can include a trajectory generator configured to evaluate one or more potential trajectories for the autonomous vehicle and to select a target trajectory based at least in part on the cost volume generate by the backbone network.