Vehicle Motion Planning with Natural Language Action Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current motion planning systems for autonomous vehicles lack the capability to interpret natural language commands, such as from traffic signs or voice commands, and integrate this interpretation with the selection of driving actions.
Innovation Solution
A system that analyzes text data representing natural language commands to generate rules with condition and result portions, and uses these rules to create a driving action mask that adjusts the probability of selecting potential driving actions, thereby aiding in motion planning decisions using reinforcement learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If motion planning systems use traditional path optimization algorithms, then computational efficiency is maintained, but the system cannot interpret natural language commands from traffic signs or voice commands
Solution Approach 1:
The patent introduces a natural language interpretation module as an intermediary between the environment (traffic signs, voice commands) and the motion planning system. This module translates natural language commands into structured rules with condition criteria and result portions, which then integrate with the existing motion planning pipeline without requiring complete system redesign.
Solution Approach 2:
The system segments the motion planning process into distinct modules: natural language command interpretation, rule generation, contradiction detection, and action selection. This segmentation allows each module to specialize in specific functions while maintaining overall system efficiency and interpretability.
2Measurement precision
If the system generates detailed rules for each natural language command, then command interpretation accuracy improves, but processing time increases
Solution Approach 1:
The system performs preliminary processing by pre-defining a vocabulary of driving actions and their associated contradiction criteria. When a natural language command is received, the system quickly matches it against predefined rules rather than generating rules from scratch, significantly reducing processing time while maintaining interpretation accuracy.
Solution Approach 2:
The system changes parameters by representing natural language commands as structured rules with specific parameters (condition criteria, result portions, contradiction indicators). This parameterized representation enables efficient comparison and processing while preserving the semantic meaning and accuracy of the original commands.
3Reliability
If the driving action mask decreases probability of contradictory actions significantly, then traffic law compliance improves, but the motion planning system loses flexibility in edge cases
Solution Approach 1:
The system applies partial masking by decreasing rather than completely eliminating the probability of contradictory actions. This allows the motion planning system to maintain traffic law compliance as the dominant behavior while retaining some probability for edge case handling, where the reinforced learning module can potentially override the mask based on learned exceptions or novel situations.
Solution Approach 2:
The driving action mask is implemented as a dynamic mechanism that adjusts action probabilities based on the interpreted natural language commands and current context. The mask parameters can be modified by the reinforced learning module, allowing the system to adapt its compliance strictness based on learned patterns and situational context, balancing reliability and flexibility.
Data Source
AI summary
A system for motion planning for a vehicle generates a rule with an associated condition criterion portion and a result portion. For each potential driving action of a set of potential driving actions of a vehicle, the system determines whether the potential driving action contradicts the result portion of the rule. A driving action mask is generated that includes a corresponding mask parameter. If the potential driving action contradicts the result portion of the rule, then the corresponding mask parameter acts to decrease a probability of selection of that potential driving action in a motion planning decision. The condition criterion portion is evaluated to determine whether the mask is applicable and if the mask is applicable, the driving action mask is used to determine a motion planning decision for the vehicle. A rule expiration criterion may also be evaluated to determine whether a rule is expired.