Autonomous Vehicle Simulation Using Safety-Critical State Editing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inefficiency of validating autonomous vehicle (AV) safety performance is hindered by the 'curse of dimensionality' and 'curse of rarity', requiring extensive real-world testing to achieve human-level safety, which is costly and time-consuming, and existing simulation methods fail to effectively simulate complex, naturalistic driving environments.
Innovation Solution
A dense deep reinforcement learning (D2RL) approach is employed to edit Markov decision processes by removing non-safety-critical states and reconnecting safety-critical states, training neural networks using only safety-critical data to create an intelligent testing environment with augmented reality, enhancing the simulation of AVs in real-world scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If extensive real-world testing is conducted to validate AV safety performance, then safety validation accuracy is improved, but testing time and cost increase significantly
Solution Approach 1:
The patent creates virtual copies of real-world driving environments, vehicles, and road users through high-fidelity simulation. These digital twins replicate physical conditions with sufficient accuracy to validate safety performance without requiring actual physical testing, thus achieving the same measurement precision with dramatically reduced time and cost.
Solution Approach 2:
The system performs preliminary validation of safety-critical scenarios in simulation before real-world testing. By pre-identifying and resolving safety issues in the virtual environment, the patent reduces the amount of time needed for actual road testing while maintaining comprehensive safety validation.
2Reliability
If traditional simulation methods are used to test AVs, then some safety validation is achieved, but the simulation cannot effectively capture complex naturalistic driving environments
Solution Approach 1:
The patent implements nested simulation where virtual vehicles contain virtual road users, which contain virtual environments. This multi-layered nesting allows each level to contribute specialized complexity - vehicle dynamics at one level, pedestrian behavior at another, and environmental conditions at a third - achieving comprehensive naturalistic driving simulation that traditional single-level simulations cannot provide.
Solution Approach 2:
The simulation system dynamically adjusts the complexity and fidelity of different environmental elements based on their relevance to safety-critical scenarios. Rather than static pre-programmed scenarios, the system generates dynamic, adaptive driving environments that can evolve during testing to match real-world naturalistic conditions.
3Loss of information
If all states in the Markov decision process are retained for training, then complete environmental information is captured, but training efficiency decreases due to non-safety-critical states
Solution Approach 1:
The patent extracts and removes non-safety-critical states from the Markov decision process training data, retaining only those states that are relevant to safety validation. This extraction process eliminates redundant information that would otherwise slow training convergence while preserving all necessary environmental information for accurate safety assessment.
Solution Approach 2:
The system applies different quality levels of state representation to different parts of the state space. Safety-critical states are represented with high fidelity and detailed information, while non-critical states are either simplified or removed entirely. This local differentiation optimizes training efficiency by focusing computational resources on the most important states.
Data Source
AI summary
A system and method for safety testing a host autonomous vehicle (AV). This method includes: generating a trained machine learning (ML) agent and testing the host AV in an environment that includes one or more background vehicles configured to operate according to the trained ML agent. The ML agent is generated by: (i) obtaining a testing state model having non-safety-critical states and safety-critical states, (ii) editing the testing state model to obtain an edited testing state model that omits data concerning the non-safety-critical states, and (iii) training a ML agent using the edited state testing model so as to generate the trained ML agent.