Avular is working on a product that uses radar and stereovision technology for the autonomous navigation of mobile robots and drones. Before our clients can start using this navigation solution, we need to prove that its software is safe. This is challenging, as no functional safety norm exists for combining these sensor technology through an AI solution. Avular has used the RobMoSys approach in software development in the Sterras project, which was completed in 2020.

The solution

Avular’s new navigation solution relies on the combination of radar data with stereovision data. As these technologies are complementary, they are the perfect match to generate reliable data in a large variety of circumstances, allowing for autonomous navigation and SLAM (simultaneous navigation and mapping). Thanks to the low hardware costs of this technology, this navigation solution will have a way lower price than existing LiDar-based solutions. On top of that, our new product can be used in bad weather conditions and dusty areas, which is not possible with LiDar technology. Based on viability study into this product, Avular sees a large application potential for this product, that has the potential to disrupt the mobile robotics market.

The first step in the project was to establish a machine-learning solution that combines radar and stereovision data, and is conformant to the RobMoSys principles of software design. A variety of ML solutions have been studied and AnyNet has been selected to estimate the depth map from stereovision cameras, as well as integrate the radar data. The end result is a system that is more resilient to visual disturbances and has a higher frame-rate than depth estimation through visual sensors alone.

After establishing a working prototype, the RobMoSys principles have been applied on the system. Formal models were created for the roles of individual components in a real-life system. Analysis showed is that, while it is not trivial to split an integrated ML system into individual components at RobMoSys framework level, through uniform data representation it is possible to add an independent health monitor to this system. This was demonstrated by integrating a health-monitoring component that analysed the AnyNet depth-map output through independent performance indicators on a robot moving through an indoor environment.

To relate the RobMoSys methodology to contemporary safety regulations for the depth-perception system, Avular has studied the generic IEC 61508 safety standard. We have found that the layered modelling approach relates closely to distinct phases in the V-model that is the basis of a E/E/PE Safety System design. In future efforts by the community we recommend that efforts are made to invest in the creation of safety standards that relate to the high-level behavioural models described in RobMoSys.

Subsidy: RobMoSys (Horizon2020)