By: Alex Thompson, Sarah Johnson, Mark Lee, Emily Chen, and David Kim
Date: July 14, 2025
Category: Amazon Bedrock, Artificial Intelligence, Automotive, Intermediate (200)
In today’s automotive landscape, the collection of vehicle data is essential for original equipment manufacturers (OEMs) seeking to enhance product innovation and performance while offering new value-added services. With the rising digitalization of vehicle systems and the adoption of software-configurable functions, OEMs are now able to efficiently integrate new features and capabilities. Sonatus’s Collector AI and Automator AI solutions tackle these two critical aspects of the shift toward Software-Defined Vehicles (SDVs) in the automotive sector.
Collector AI simplifies the use of data throughout the vehicle lifecycle by allowing the creation of data collection policies without any alterations to vehicle electronics or embedded code. However, OEM engineers and other vehicle data users often find themselves overwhelmed by the multitude of vehicle signals necessary to address their specific use cases. Similarly, Automator AI’s no-code approach to automating vehicle functions through intuitive if-then scripted workflows can pose challenges, particularly for OEM users who may lack expertise in vehicle events and signals needed for their desired automated actions.
To overcome these hurdles, Sonatus collaborated with the AWS Generative AI Innovation Center to create a natural language interface that generates data collection and automation policies using generative AI. This advancement aims to condense the policy generation timeline from days to mere minutes, making it accessible to both engineers and non-technical users alike.
In this article, we will discuss how we developed this innovative system leveraging Sonatus’s Collector AI and Amazon Bedrock. We will cover the foundational concepts, challenges faced, and the overarching architecture of the solution.
Collector AI and Automator AI
Sonatus has crafted an advanced vehicle data collection and automation workflow tool encompassing two primary components:
- Collector AI – Captures and relays precise vehicle data based on customizable trigger events.
- Automator AI – Implements automated actions within the vehicle based on analyzed data and specified trigger conditions.
Currently, engineers must manually create data collection or automation policies, which can number in the hundreds for a single vehicle model depending on the OEM’s range of use cases. Furthermore, identifying the appropriate data to collect for a specific intent involves navigating through various layers of information and organizational complexities. Our objective was to create a more intelligent and intuitive method to:
- Generate policies from natural language input
- Dramatically cut policy creation time from days to minutes
- Maintain control over intermediate steps in the generation process
- Enable non-engineers, such as vehicle product owners and planners, to participate in policy creation
- Incorporate a human-in-the-loop review system for both existing and newly crafted policies
Key Challenges
During implementation, we faced several challenges:
- Complex Event Structures – Different vehicle models and policy entities utilize varied representations and formats, necessitating flexible policy generation.
- Labeled Data Limitations – There is a scarcity of labeled data to accurately map natural language inputs to the desired policies.
- Format Translation – The solution must effectively handle diverse data formats and schemas across different customers and vehicle models.
- Quality Assurance – Ensuring the accuracy and consistency of generated policies is paramount.
- Explainability – Providing clear explanations of how policies are generated fosters trust among users.
Success Metrics
We established key metrics to evaluate the success of our solution:
- Business Metrics:
- Decreased policy generation time
- Increased number of policies for each customer
- Broadened user base for policy creation
- Technical Metrics:
- Accuracy of generated policies
- Quality of outcomes for modified prompts
- Operational Metrics:
- Reduced effort and turnaround time for policy generation compared to the manual process
- Successful integration with existing systems
Solution Overview
The collaboration between the Sonatus Advanced Technology team and the Generative AI Innovation Center produced an automated policy generation system, depicted in the accompanying diagram.
This system consists of a series of large language models (LLMs) performing individual tasks, including entity extraction, signal translation, and signal parametrization.
Entity Extraction
A fully generated vehicle policy comprises multiple components, derived from a single user statement. These include triggers and target data for collector policies, as well as triggers, actions, and tasks for automator policies. The user’s statement undergoes a breakdown into its entities through specific rules and steps:
- Few-shot examples are provided for each entity.
- Trigger outputs must contain the relevant signal value and comparison operator information.
- For instance, a query like: “Create an automation policy that locks the doors when the car is in motion” would yield: vehicle speed above 0, vehicle signal
- Triggers and actions are subsequently verified using a classification prompt.
- For Automator AI, triggers and actions must correspond with their associated tasks.
The final output of this process is an intermediate structured XML representation of the user query in natural language.
Signal Translation and Parametrization
To derive the final JSON policy structure from the intermediate XML output, the correct signals must be identified, and the signal parameters generated. This information is then combined to adhere to the application’s expected JSON schema.
The preferred output signal format at this stage is the Vehicle Signal Specification (VSS), an industry-standard specification led by COVESA. VSS defines vehicle signal naming conventions and strategies, making vehicle signals more descriptive and understandable compared to their physical Control Area Network (CAN) signal counterparts. This descriptive quality is vital in the generative AI process, as it requires clear signal names and the understanding of their meanings.
The VSS signals, along with their definitions and necessary metadata, are integrated into a vector index. For each XML structure needing a vehicle signal lookup, the process of signal translation includes several steps:
- Available signal data is preprocessed and stored in a vector database.
- Each XML representation—triggers, actions, and data—is converted into corresponding embeddings. At times, the XML phrasing may introduce minor errors, which can be rectified during the translation.
For further insights, you can check out this related blog post, and for a deeper understanding of the subject, refer to this authoritative source. Additionally, if you’re interested in enhancing your knowledge, this resource on learning and development is an excellent option.
Leave a Reply