It’s a complex ever-changing world out there and your autonomous vehicle needs to navigate it safely. But getting your AI perception system on the road can be delayed by an expensive cycle of testing and data collection.
aidkit breaks the cycle and gets your processes flowing through comprehensive testing and guidance for lean data collection and labelling.
No matter the model, scene, or ODD, aidkit ’til you make it.
aidkit is the toolkit to test the perception functions of ADAS/ADS and ensure they are safe for deployment.
By testing your ML model for robustness against a range of altered scenes, aidkit ensures reliable performance in your ODD.
Where performance falters, aidkit guides your data campaign so you know how to improve.
aidkit integrates into existing workflows and reduces time to deployment though useful features such as scaling, data lineage, and automatic retraining.
Deploy a safer perception component with the empirical evidence to back up your safety claims to safety managers and external regulators.
aidkit generates a wide range of corrupted images for performance- and worst-case analysis to ensure the robust handling of common cases and identify edge and corner cases within and around your ODD.
Testing with aidkit allows you to know exactly what data to collect and label to meet your performance shortfalls so you can tailor your data campaign to fill gaps without expensive, extraneous collection and labeling.
aidkit provides you with explainable, evidence-based risk scores containing impact and probability of occurrence for every tested ODD scenario and a complete record of how you got your results.
aidkit: trust in your AI begins with testing you can trust
By augmenting data with perturbation-based image operations, aidkit tests the performance of perception models under the various conditions that are possible in an ODD but not easily captured in fleet data. This technique fills the gap between reality and simulation. Along with real-world data for initial training and simulation for later testing, augmented data should be a pillar in your methodological triangulation: the key to a strong scientific basis and confidence in your AI model.
aidkit makes these complex ideas and tasks easy to implement and understand, allowing you to translate the requirements of current and upcoming safety standards into technical pass/fail criteria and then clearly communicate you've met the criteria.
aidkit can test segmentation, detection, or classification models from PyTorch to Tensorflow
Choose your metrics, set the parameters, and customize the corruptions for your tests
Have a coffee break, but with scalability and interim monitoring don’t get too comfortable!
Get your risk scores reports and other statistics to present your safety case or plan how to improve
Automatically use artifacts for further testing and see where you need more data
Driving safe perception
Don’t just manage risks. Mitigate your perception system’s chances of encountering unknown situations and enhance its resiliance.
© neurocat GmbH