Unmanned Systems Technology Dec/Jan 2020 | Phoenix UAS | Sonar focus | Construction insight | InterGeo 2019 | Supacat ATMP | Adelan fuel cell | Oregon tour | DSEI 2019 | Copperstone Helix | Power management focus

98 PS | Deep learning for autonomous cars A chieving the highest levels of autonomy for ground vehicles is proving harder than a naturally optimistic industry has been hoping, and much of the hype from the car industry in particular is now deflating (writes Peter Donaldson). Road systems are complex and can be chaotic, even though they are structured and governed by rules. The problem is so hard that established AI approaches to ground vehicle autonomy may have to give way to new ones. Road networks have markings for lanes and junctions, road signs, traffic lights, designated crossings, buildings and landmarks, street lights, highway codes and traffic laws and traffic officers to enforce them. Even so, producing vehicles that can drive themselves safely in these environments seems further off now than it did just a few years ago. Current software finds it difficult to predict the actions of other road users and pedestrians, and to cope with bad weather and other real-world vagaries. That makes SAE Level 5 autonomy – the ability to drive anywhere under all conditions – very difficult to achieve. The fundamental problem seems to be that while deep learning (DL), the dominant form of machine learning applied to autonomous vehicles, does not have a high-level understanding of driving. It is a means of matching statistical characteristics with outcomes, and is at its best when presented with clearly bounded and unchanging data sets. Unfortunately, the real world will never present the algorithms with such clean and convenient data. However, AI research company Wayve may have found a way through this problem with a new take on end-to-end DL using a combination of reinforcement learning, supervised learning and self- supervised learning plus computer vision. Last year it demonstrated what it calls the first end-to-end DL driving system with full control of the vehicle and following a user-defined route. The focus was on copying human driving, using conditional imitation learning to train a neural network. With 30 hours of human driving data it trained a machine learning model to copy the behaviour, learning a simple motion plan directly from camera images. It then carried out 35 km of autonomous testing on two urban routes that did not feature in the training data, during which it also forced the test car to interact with a second car that was acting as a mobile hazard. From this, Wayve concluded that it is crucial to weight the data properly to avoid ‘naive’ training that in this case might result in a model that is only good at driving in a straight line and sitting in traffic. It also learned the importance of computer vision that, for example, can cope with different lighting and weather conditions to enable its models to generalise about the real world. To tackle this it set the computer vision tasks including semantic segmentation to tell the model what is around it, monocular depth estimation to work out the geometry of the scene, and optical flow estimation to calculate how different parts of the scene are moving. Finally, Wayve found that using multiple sources of information helped it to develop a useful driving policy from only 30 hours of data. These included imitating expert human drivers, labelled semantic computer vision data, safety driver intervention, corrective action and simulation. Others included geometry, dynamics, motion and future prediction. It’s too early to say whether it has found a silver bullet with this approach, but the results look promising. Now, here’s a thing “ ” The company found that using multiple sources of information helped develop a useful driving policy from only 30 hours of data December/January 2020 | Unmanned Systems Technology