Unmanned Systems Technology 013 | AutonomouStuff Lincoln MKZ | AI systems | Unmanned Underwater Vehicles | Cosworth AG2 UAV twin | AceCore Neo | Maintenance | IDEX 2017 Show report

Static versus dynamic learning Capturing data from all the autonomous vehicles on the road and using that data to improve the network is eminently sensible, and in fact is already being done. Having a learning network within a vehicle is outside the projected performance of hardware for several years to come, so at the moment other approaches are being considered. These involve sending data such as images and speeds back to a central learning network in the cloud. This large learning network would then generate new weightings and possibly new configurations for the neural network in the vehicle. However, it is not easy to carry out the runtime update on the vehicle, nor is it necessarily desirable. Instead, data can be fed back to the training engine to improve the accuracy of the requirements leads to path planning. This looks at where the vehicle intends to travel and how that impacts on the other agents in its environment. For example, a car driving past a pedestrian will build a model of what the pedestrian intends to do, and plan a route accordingly. The planner may even produce multiple models with different routes and speeds, then choose the one with the lowest acceptable level of risk. That might not be the lowest risk option though, as it could mean the car doesn’t actually move. Path planning also uses inverse reinforcement learning, looking at what the agents do and working backwards to establish why they did that. This is a highly complex modelling environment. One of the challenges of deep learning is that it is a probabilistic system – a slight difference in the input may produce a very different outcome, and there is no way to tell why that happened. Being able to explain how a decision is reached is important for a safety-critical design process, and that is driving certain architectural choices. Rather than driving millions of miles with an AI system to show that it is accurate, the system is broken down into separate modules that can each be tested to demonstrate repeatable, reliable and consistent results. What these modules are is a matter of the evolving system design. For example, they can be individual neural nets for different types of data, different agent modelling modules and path planning analysis and execution. For more complex urban driving it is difficult to be confident that the end-to-end network has been taught everything it might need to know, so the modular approach allows analysis of the separate elements, such as the cars and pedestrians it ‘sees’ and the resulting decisions on the chosen path.