Autonomous Vehicle Motion Planning With Shared-Backbone Trajectory Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing motion planning architectures for autonomous vehicles suffer from high computational expense, energy usage, and latency due to iterative optimization techniques that limit the scope of candidate trajectories without full knowledge of their implementation, leading to imperfect information and suboptimal decision-making.
Innovation Solution
A multistage machine-learned model architecture for autonomous vehicle motion planning, comprising a first stage for generating candidate trajectories with high recall and a second stage for ranking them with high precision, leveraging a shared backbone architecture to process context data and evaluate interactions with the environment, allowing parallelizable batch generation and scoring of trajectories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative optimization techniques are used to evaluate candidate trajectories, then measurement precision of trajectory selection is improved, but loss of time and energy consumption increase significantly
Solution Approach 1:
The system performs preliminary action by generating a large set of candidate trajectories with high recall in the first stage using a machine-learned model, ensuring that the optimal trajectory is included in the candidate set before the ranking stage. This preliminary generation with full scene context access prevents the need for iterative optimization later, as the candidates are already well-prepared for evaluation.
Solution Approach 2:
The motion planning architecture is segmented into two distinct stages: a first stage for generating candidate trajectories with high recall, and a second stage for ranking them with high precision. This segmentation allows each stage to be optimized independently - the first stage focuses on comprehensive candidate generation without iterative optimization, while the second stage performs efficient ranking on a reduced set, thereby reducing overall computational latency and energy consumption.
2Device complexity
If the scope of candidate trajectories is limited without full knowledge of implementation, then device complexity is reduced, but loss of information occurs leading to suboptimal decision-making
Solution Approach 1:
The system performs preliminary action by generating candidate trajectories with high recall that closely represent the actual movements to be executed by the autonomous vehicle. This preliminary representation ensures that full scene context information is preserved in the candidate trajectories before ranking, preventing information loss while avoiding the need for complex iterative optimization to discover better candidates.
Solution Approach 2:
The system generates a large number of candidate trajectories in the first stage that serve as disposable candidates for evaluation in the second stage. These candidates are generated efficiently without iterative optimization, and while individual candidates may not be perfect, the high-recall generation ensures the optimal trajectory is included. The ranking stage then efficiently selects the best candidate, avoiding the need for complex iterative refinement of each candidate.
3Manufacturing precision
If high-recall configuration is used to increase likelihood of including good trajectories, then manufacturing precision of trajectory selection is improved, but use of energy increases due to processing more candidates
Solution Approach 1:
The energy-intensive task of trajectory generation and evaluation is segmented into two stages. The first stage uses a machine-learned model to generate candidates with high recall efficiently, leveraging the model's ability to process full scene context in one pass. The second stage ranks the candidates with high precision using a more computationally efficient process that operates on the reduced candidate set, rather than performing iterative optimization on all possible trajectories. This segmentation reduces overall energy consumption while maintaining high trajectory selection quality.
Solution Approach 2:
The system performs preliminary action by generating a comprehensive set of candidate trajectories with high recall using the trained machine-learned model, which has already learned efficient patterns from training data. This preliminary generation avoids the need for energy-intensive iterative optimization during deployment, as the model directly produces high-quality candidates based on learned representations of scene context and trajectory relationships.
Data Source
AI summary
The present disclosure provides an example method that includes: (a) obtaining context data descriptive of an environment surrounding an autonomous vehicle, the context data based on map data and perception data; (b) generating, by a proposer and based on the context data: (i) a plurality of candidate trajectories, and (ii) a plurality of actor forecasts for a plurality of actors in the environment; (c) generating, by a ranker and based on the context data, the plurality of candidate trajectories, and the plurality of actor forecasts, a ranking of the plurality of candidate trajectories; and (d) controlling a motion of the autonomous vehicle based on a candidate trajectory selected based on the ranking of the plurality of candidate trajectories, wherein the proposer comprises a first machine-learned model and the ranker comprises a second machine-learned model, and wherein the first machine-learned model and the second machine-learned model use a common backbone architecture.