Motion Planning with Predicted Occupancy Grids for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous driving systems face challenges in propagating uncertainty and sensitivity between perception and planning modules, and existing path planning strategies do not effectively utilize data from human drivers, leading to suboptimal performance in dynamic and unstructured environments.
Innovation Solution
A machine-learning based system for end-to-end path planning that integrates perception, prediction, and planning modules, using a neural network to generate trajectories from sensor data, including LiDAR, and predicts Occupancy Grid Maps (OGMs) to adapt to dynamic environments, allowing for modular or end-to-end training and error propagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If modules are isolated from each other, then each module can be developed independently, but uncertainty does not propagate properly through the system
Solution Approach 1:
The system divides the autonomous driving function into distinct modules (perception module, prediction module, planning module) that can be developed and trained independently. Each module processes specific aspects of the driving task while maintaining defined interfaces for data exchange.
Solution Approach 2:
The system implements feedback mechanisms where uncertainty information from each module is propagated to subsequent modules. The planning module receives not only trajectory predictions but also uncertainty measurements from the prediction module, allowing it to adjust its planning accordingly.
2Ease of manufacture
If path planning uses predefined costs by designer, then engineering control is improved, but a lot of engineering work is required to become sufficiently robust
Solution Approach 1:
The system enables the planning module to automatically learn optimal cost functions and planning strategies from human driving data. Instead of requiring extensive manual engineering of cost parameters, the system self-adjusts its planning behavior through machine learning, reducing the time and expertise required for engineering tuning.
3Extent of automation
If simulator is used to generate training data, then data generation is automated, but training a simulator is complex and difficult to debug
Solution Approach 1:
The system uses real human driving data as a copy of actual driving behavior to train the planning module. Instead of relying on simulated data that requires complex simulator training, the system directly learns from recorded human driving trajectories, simplifying the training process while maintaining high fidelity.
4Manufacturing precision
If path planning is performed based on predefined costs, then design control is improved, but robustness requires significant engineering effort
Solution Approach 1:
The system transitions from static, predefined cost functions to dynamic, learned cost functions that adapt to different driving scenarios. The planning module learns to adjust its cost parameters based on the specific situation, improving robustness across diverse environments while maintaining design control through the structured learning framework.
Data Source
AI summary
A system and method for path and/or motion planning and for training such a system are described. In one aspect, the method comprises generating a sequence of predicted occupancy grid maps (OGMs) for T−T1 time steps based on a sequence of OGMs for 0−T1 time steps, a reference map of an environment in which an autonomous vehicle is operating, and a trajectory. A cost volume is generated for the sequence of predicted OGMs. The cost volume comprises a plurality of cost maps for T−T1 time steps. Each cost map corresponds to a predicted OGM in the sequence of predicted OGMs and has the same dimensions as the corresponding predicted OGM. Each cost map comprises a plurality of cells. Each cell in the cost map represents a cost of the cell in corresponding predicted OGM being occupied in accordance with a policy defined by a policy function.