ML Maneuver Output Biasing for Risk-Adaptive Aerial Vehicle Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ML algorithms in safety-critical applications like aerial vehicle flight control systems do not allow for alterations in output decisions based on acceptable risk levels, limiting the ability to adjust maneuver recommendations to increase or decrease structural risk as needed for specific missions or conditions.
Innovation Solution
Implementing an output biasing module that tunes ML maneuver recommendation engines by adjusting maneuver confidence thresholds using risk tuning parameters, allowing for increased or decreased structural risk based on mission parameters and environmental conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ML algorithms use fixed output decisions based on training data, then prediction consistency is improved, but adaptability to different risk levels deteriorates
Solution Approach 1:
The patent implements dynamic adjustability of output decisions by introducing a mechanism that modifies ML algorithm outputs based on runtime risk level parameters. The system transitions from static fixed predictions to dynamic adaptive predictions by applying risk level-based modifications to maneuver recommendations, allowing the same ML model to serve multiple risk tolerance scenarios without retraining.
Solution Approach 2:
The patent changes the parameter of output decision confidence thresholds by introducing risk level parameters that adjust the stringency of maneuver recommendations. By modifying the confidence threshold parameter based on risk level, the system enables the ML algorithm to produce different output decisions for the same input data depending on the acceptable risk level, thus achieving adaptability while maintaining prediction consistency through the underlying trained model.
2Reliability
If maneuver recommendations follow strict safety rules, then structural safety is improved, but mission completion capability deteriorates
Solution Approach 1:
The patent implements dynamic adjustment of safety rule enforcement by introducing risk level parameters that modulate the strictness of maneuver recommendations. At lower risk levels, the system applies stricter safety constraints to maximize structural safety. At higher risk levels, the system relaxes constraints to allow more aggressive maneuvers that may be necessary for mission completion, thus dynamically balancing safety and productivity based on operational context.
Solution Approach 2:
The patent changes the parameter of confidence thresholds for maneuver recommendations based on risk level settings. By adjusting this parameter, the system controls the trade-off between safety and mission completion: higher confidence thresholds produce more conservative, safety-oriented recommendations, while lower thresholds enable bolder maneuvers that increase mission completion likelihood but also structural risk.
3Object-affected harmful factors
If ML algorithms avoid high-risk maneuvers, then structural risk is reduced, but ability to reach safe locations deteriorates
Solution Approach 1:
The patent changes the risk tolerance parameter of the ML algorithm outputs by introducing risk level adjustments. This allows the system to modulate the aggressiveness of maneuver recommendations: at risk-averse settings, the algorithm prioritizes avoiding high-risk maneuvers; at risk-tolerant settings, it accepts higher-risk maneuvers when they provide the only path to reaching safe locations, thus adapting the risk-reward balance to mission requirements.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In some examples, systems and methods are described for output biasing maneuvers recommendations provided by at least one machine learning maneuver recommendation (MLM) engine executing on an aerial vehicle. In some examples, output biasing data can be received that includes at least one risk tuning parameter that can influence which of the maneuver recommendations are selected by a maneuver decision engine executing on the aerial vehicle based on a maneuver confidence threshold for implementation by the aerial vehicle. The maneuver confidence threshold can be updated based on the at least one risk tuning parameter to provide an updated maneuver confidence threshold for the output biasing of the maneuvers recommendation provided by the at least one MLM engine. Vehicle command data for implementing a given maneuver recommendation can be outputted based on an evaluation of the updated maneuver confidence threshold.