Autonomous Vehicle Maneuver Selection With Safe Action Masking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing autonomous vehicle control systems using reinforcement learning (RL) lack safety guarantees, leading to potential unsafe decisions due to the prioritization of efficiency and comfort over safety during training, which complicates training and obscures the understanding of action effects.

Innovation Solution

A safety framework is integrated with a neural network architecture to predefine safe and legal actions, limiting the action space and ensuring only those actions are considered, using mechanisms like dot products and softmax activation to filter out unsafe maneuvers, thereby providing transparent and efficient training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If reinforcement learning prioritizes efficiency and comfort during training, then training convergence speed improves, but safety of vehicle decisions deteriorates

Engineering Contradiction:
Improvetraining convergence speedVSAvoidsafety of vehicle decisions
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary action by pre-defining the action space to include only safe and legal maneuvers before training begins. The safety framework and neural network architecture are configured upfront to filter out unsafe actions, ensuring that the RL agent only learns from safe experiences. This resolves the contradiction by establishing safety constraints before efficiency optimization occurs during training.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If the action space is not limited, then the agent can explore more possibilities, but transparency of action understanding deteriorates

Engineering Contradiction:
Improveexploration capabilityVSAvoidtransparency of action understanding
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent applies local quality by making the action space heterogeneous with different properties - some actions are marked as safe/legal while others are unsafe/illegal. The neural network outputs probabilities for each action type, and the safety framework selectively enables/disables specific actions based on their safety classification. This allows the agent to explore within safe boundaries while maintaining transparency about which actions are considered safe versus unsafe.

Inventive Principle:
Principle #3Local quality

3Reliability

If unsafe actions are filtered out after execution, then safety is maintained, but training efficiency deteriorates

Engineering Contradiction:
Improvesafety guaranteeVSAvoidtraining efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies preliminary action by configuring the neural network architecture and safety framework before training to prevent unsafe actions from being selected in the first place. The action space is pre-filtered to only include safe maneuvers, and the RL agent receives feedback only on safe actions. This eliminates the need for post-execution filtering and avoids wasted training iterations on unsafe actions, thereby maintaining safety while improving training efficiency.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4001041B1Method and system for determining a maneuver to be executed by an autonomous vehicle
Publication Date: 2026.03.18 APTIV TECHNOLOGIES AG
  • EP4001041B1 patent drawingFigure 1
  • EP4001041B1 patent drawingFigure 2
  • EP4001041B1 patent drawingFigure 3

AI summary

A computer implemented method for determining a maneuver to be executed by an autonomous vehicle comprises the following steps carried out by computer hardware components: determining a plurality of potential maneuvers; determining a subset of the potential maneuvers based on a pre-determined condition; determining a mask based on the subset; determining an output of a fully connected layer of a neural network; determining a dot product of the output of the fully connected layer and the mask; and determining the maneuver to be executed based on the dot product.