A neural network structure may provide the control – and thus the safety – people are looking for in driverless cars, if a successful trial in the US is any guide.
Engineers at Stanford University certainly believe it warrants further investigation. They used such an approach to create an autonomous vehicle (AV) that could more than match the speed and precision of a champion amateur race car driver.
As Nathan Spielberg and colleagues explain in a paper in the journal Science Robotics, to dodge obstacles better than an average driver, an AV will need to drive at the limit of friction – just before the tyres stop holding the car to the road and it begins to swerve and spin out.
Being able to operate effectively at these limits, when most drivers lose control, is essential for avoiding high-speed accidents or when driving in slippery conditions.
The issue, they say, is that most current autonomous driving systems rely on estimates of the road-tyre friction to inform speed and steering control, but available methods for doing this are not very accurate, particularly when friction changes quickly with road conditions.
Their answer was to create a “feedforward-feedback” system that uses mathematical modelling of the physical forces acting on the vehicle to predict the optimal steering angle without taking estimates of the road friction.
A car with this system was able to navigate a course as quickly as the race car driver, and the researchers were able to improve performance further by replacing the physics-based model with a neural network model trained by driving the vehicle in various conditions.
“The neural network achieved better performance than the physical model when implemented in the same feedforward-feedback control architecture on an experimental vehicle,” they write.
“More notably, when trained on a combination of data from dry roads and snow, the model was able to make appropriate predictions for the road surface on which the vehicle was travelling without the need for explicit road friction estimation.”
Spielberg and colleagues used the states and inputs of the physics-based model to develop a two-layer feedforward neural network capable of learning vehicle dynamic behaviour on a range of different surfaces.
The network involves a combination of current measurements and history information from three previous time steps. The history information enables the network to provide predictions of behaviour at different friction levels without the need for an explicit friction estimation scheme.
“When trained on a combination of high- and low-friction data, the model made predictions appropriate to the surface described by the history information,” they say.
“By foregoing the step of friction estimation, the neural network with history information fused estimation and prediction capabilities, simplifying the task of vehicle control. This additional functionality did not come at the expense of performance.”