Autonomous Vehicle Control Model Refinement for Safety-Progress Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learned models for autonomous vehicles often improve some performance metrics while regressing others, failing to meet stringent safety and efficacy requirements for autonomous control.
Innovation Solution
A refinement training process that uses vehicle and driver trajectories to determine a preferred trajectory, adjusting the machine-learned model to reduce loss and improve safety, progress, and comfort metrics by targeting scenarios where the model performed poorly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine-learned models are trained to improve safety metrics, then collision avoidance improves, but other performance metrics such as progress or comfort may regress
Solution Approach 1:
The patent segments the training process into multiple distinct phases: pre-training phase, refinement training phase, and evaluation phase. Each phase targets specific performance metrics with dedicated training data and objectives, allowing simultaneous optimization of safety, progress, and comfort without mutual interference
Solution Approach 2:
The patent changes training parameters dynamically by adjusting the composition of training scenarios based on performance metric thresholds. When safety metrics fall below thresholds, the training data is enriched with safety-critical scenarios; when progress metrics regress, the training includes more efficiency-focused scenarios, allowing flexible optimization of multiple metrics
2Ease of operation
If machine-learned models are trained to improve comfort metrics, then jerky behavior reduces, but safety or progress performance may regress
Solution Approach 1:
The training process is segmented into distinct phases where comfort optimization occurs in the refinement training phase using comfort-focused scenarios, while safety is maintained through the pre-training phase and continuous evaluation, preventing regression in safety performance
Solution Approach 2:
The patent implements feedback loops where performance metrics are continuously evaluated after training iterations. If comfort metrics improve without safety regression, the training continues; if safety regresses, the training parameters are adjusted to restore safety performance before further comfort optimization
3Reliability
If extensive training data and scenarios are used to improve all performance metrics, then model performance improves, but computational power and time consumption increase
Solution Approach 1:
The patent applies partial action by using threshold-based scenario selection: only scenarios that cause performance metric regresses are included in re-training, rather than using all possible scenarios. This reduces computational load while maintaining performance improvement
Solution Approach 2:
The patent performs preliminary evaluation of training scenarios to identify those most likely to improve performance metrics before actual training. Scenarios are pre-filtered based on their potential impact on safety, progress, and comfort metrics, reducing the computational burden during actual training phases
Data Source
AI summary
A machine-learned model that uses sensor and/or perception data to directly determine controls for operating an autonomous vehicle may be trained by identifying a preferred trajectory between a human-driven and vehicle-controlled trajectory, and using a first loss determined between the vehicle-controlled trajectory and the path the autonomous vehicle ultimately ended up taking in a scenario and a second loss determined between the vehicle-controlled trajectory and the human-driven trajectory to refine the machine-learned model. The machine-learned model may additionally or alternatively be refined by a learned reward model constructed by replacing one or more output heads of the machine-learned model with a regression head that is trained using performance metrics determined for the vehicle-controlled trajectory.