On the road towards fully autonomous vehicles, there will be a period of transition where people will be required to retake control from autonomous systems at certain times, as is currently the case with Tesla's Autopilot. Researchers at Stanford University have been looking specifically at this handover period and found that the transition from autonomous driving to manual control requires some adjustment by the human involved, which could open the window to accidents.

The Stanford study involved 22 drivers on a closed track driving study-designed autonomous cars. The cars took the drivers around the track and then had the drivers take over the controls at various points to measure reactions to changing conditions. The measurements included steering adaptation, speed changes, and varied movements, such as turning and lane changing.

Even with advance warning and full knowledge of what changes were coming, drivers still showed significant transitional adjustments that indicate that a window of time will exist where the vehicle will be particularly prone to accidents.

The study participants drove a 15-second course with a straightaway and lane change. The autonomous car was capable of driving the entire course on its own, including looping back to the beginning, and several runs were made with each driver taking over control during the course. Reactions when driving without autonomy, with autonomy, and during transitions were measured over a total of 14 runs per driver. For 10 of the runs, steering conditions were modified to represent changes in speed and steering that occur under autonomous control.

Lead author of the study, Holly Russell, a former graduate student in the Dynamic Design Lab(Credit: Steve Castillo)

Steering ratios were the primary means of speed change representation. A change from a 15:1 turn ratio to a 2:1 ratio represented the change from low-speed to freeway-speed driving, for example, making the car turn more sharply to simulate highway speeds.

Although drivers knew about these changes in advance and had some time to prepare themselves, their maneuvering still differed significantly when taking over from autonomous driving than it did when piloting the car entirely on their own, with the drivers wobbling the wheel to account for over- and under-steering. So even in a best-case scenario, the takeover from autonomous to manual driving results in a period of increased risk of accident.

"Even knowing about the change, being able to make a plan and do some explicit motor planning for how to compensate, you still saw a very different steering behavior and compromised performance," said Lene Harbott, co-author of the study.

Neuroscience explains this phenomenon as the difference between explicit and implicit learning. Implicit motor control requires actually performing a task to learn it and cannot be accurately replicated by explicit awareness and planning.

None of the drivers in the Stanford study drove off course as a result of the transition from autonomous driving to manual control, but the study represents a best-case scenario and shows that even then, there is a period of altered steering behavior. That period, the study authors say, could mean accidents on the road when autonomous cars become more common.

The researchers also point out that the test vehicle used is their own design and does not represent any specific manufacturer's commercial system under development. Researchers say that each of those commercial systems should undergo similar tests to measure how that particular autonomous system transitions to manual control in varied conditions.

"If someone is designing a method for automated vehicle handover, there will need to be detailed research on that specific method," says Harbott. "This study is tip of an iceberg."

Source: Stanford University

View gallery - 2 images