A driverless car can make high-speed turns using a machine learning algorithm that studied footage of similar manouevres.
The type of artificial intelligence is called a neural network and is loosely based on how our brains work.
More than 200,000 motion samples taken from test drives on an icy track near the Arctic circle to train the self-driving system not to spin out of control.
It observed its motion from previous fractions of a second to adjust its steering to provide accurate motion predictions on different road surfaces.
The team from Stanford University in California equipped a Volkswagen GTI with the algorithm and tested it on an oval-shaped race track.
Driving as fast as physically possible and having learned from watching previous tests, the car adjusted its steering and acceleration to turn successfully.
For autonomous vehicles to operate safely, they need control systems that can rapidly brake, accelerate or steer in critical situations.
This allows them to drive safely at the limits of friction - just before the tyres stop gripping the road and the car spins out.
'With the techniques available today, you often have to choose between data-driven methods and approaches grounded in fundamental physics,' said J. Christian Gerdes, a professor of mechanical engineering and senior author of the paper.
Professor Gerdes said his system could help in emergency situations, where sudden swerves are needed.
The results were encouraging, but the researchers stress that their neural network system does not perform well in conditions outside the ones it has experienced.
Professor Gerdes said that one challenge of the neural network is a lack of insight into how it works.
'If you give it a set of conditions it hasn't seen before, it may extrapolate in ways that are completely wrong,' resulting in potentially dangerous steering controls, he said.
The team are continuing to develop the system and vehicles that have seen thousands of times more interactions than we have, we can hopefully make them safer.
The researchers are now building safety features into the system to check its decisions are reasonable.
The research was published in Science Robotics.
