In the autonomous vehicle industry, understanding which elements of a scene actually influence the model's planning decisions is a critical challenge for safety and trust. Techniques such as counterfactual analysis allow isolating the causal effect of individual objects on the predicted path, an approach that researchers have explored by selectively removing objects from images using generative inpainting. This methodology, similar to the framework proposed in Counterfactual Vision Action Analysis, reveals that vehicles and pedestrians within the expected trajectory have a dominant influence, while traffic lights show a disproportionate effect relative to their size in the image. However, surprises also emerge: the model may react strongly to objects that a human driver would consider irrelevant, raising profound questions about the system’s internal representation.
From a business perspective, these findings underscore the need for explainable and auditable artificial intelligence systems. At Q2BSTUDIO, a company specialized in software and technology development, we work to help organizations of all sizes implement transparent AI solutions, whether through custom software applications for autonomous mobility or predictive analytics systems. Our team integrates counterfactual and mechanistic interpretability methodologies to audit neural network models, identifying which features –beyond human-readable objects– are actually encoded in the internal layers. This type of analysis is essential when deploying AI agents in critical environments such as autonomous driving or industrial robotics, where a misinterpretation can have severe consequences.
The combination of computer vision techniques with generative artificial intelligence allows creating counterfactual datasets, such as those generated by removing objects from real scenes via inpainting. In these scenarios, the model reveals hidden biases: for example, it may give weight to irrelevant traffic signs or ignore partially occluded pedestrians. To address this, at Q2BSTUDIO we propose solutions that integrate cloud infrastructure on AWS and Azure, enabling training and evaluation of models with large volumes of counterfactual data, ensuring scalability and performance. Our cloud AWS/Azure service provides the ideal environment to process datasets like Counter-nuScenes, where each image requires an inpainting pipeline with generative models such as Stable Diffusion or GANs, followed by statistical analysis of prediction differences.
Beyond mobility, counterfactual philosophy applies to sectors such as healthcare, finance, and logistics. When a company needs to understand why a credit scoring model rejected an application, or why a recommendation system prioritized one product over another, counterfactual analysis offers clear answers. At Q2BSTUDIO we integrate these principles into our Business Intelligence solutions with Power BI, allowing analysts to visualize not only what a model predicts, but also how the prediction would change if a specific variable were modified. This is particularly useful in regulated environments where explainability is a legal requirement, such as in cybersecurity risk assessment or AI algorithm auditing.
Cybersecurity is another area where a counterfactual lens proves powerful. By simulating attacks or removing nodes in a network, one can determine which assets are critical to system security. At Q2BSTUDIO we offer cybersecurity and pentesting services that include AI model analysis, assessing their robustness against adversarial manipulations. Additionally, our AI agent solutions based on artificial intelligence automate incident responses, learning from counterfactual patterns to improve anomaly detection.
Ultimately, the question 'What do autonomous vehicles see?' transcends the academic realm and becomes an industrial imperative. Companies that bet on transparency in their artificial intelligence systems not only build trust but also reduce operational and legal risks. At Q2BSTUDIO, we help our clients design, audit, and optimize models using counterfactual techniques, integrating AI, cloud, cybersecurity, and Business Intelligence capabilities. Whether through custom software for computer vision or cloud-based analytics platforms, our goal is to make every automated decision as understandable as it is accurate. The future of autonomous mobility depends on technology that not only works but can also explain itself. And in that task, counterfactual analysis is an indispensable tool.
