We just came across an interesting interview by Sebastian Thrun, head of Google X and mastermind behind Google’s Driverless Car. In the interview conducted by Charlie Rose in late April he talks about Google’s project Glass, Udacity and driverless cars. He sounds a cautious note on the safety of autonomous cars. While the technology is already quite safe, he is still concerned about its capability of driving millions and millions of miles without error. He suggests that current driverless cars have not quite reached the perfomance level of an attentive human driver. In another interview with WIRED he also discusses the problem hinting, that the right combination between human and computer intelligence may need to be found to ensure maximum safety.
Are we right to conclude that Google’s driverless car team is finding it difficult to reach the intended safety level for their cars? From a statistical perspective it is very hard to prove that an autonomous vehicle can perform flawlessly for millions of miles. Just driving a few million miles in test mode is not enough.
So far Google has always emphasized that their driverless cars drive on known routes for which detailed navigation and localization data is available. Relying on stored information about a route, however, can be a major source of error. Therefore Google must also be working hard to run their cars without prerecorded mapping data. It would be interesting to know their progress in this area.
An additional approach for verifying the safety of their cars would be to develop a full-fledged simulator built around a physics engine which would be connected to the sensor and actuator interfaces of Google’s driverless car (for an early sketch of this approach, see our Innovation Brief (2010)). The driverless car control unit would receive sensor inputs generated by the simulator. These sensor inputs would be updated according to the simulated movement of the car depending on the signals received by the actuators. Building such a simulator would be a major challenge because of the large amount of sensor data which it would need to generate. But Google’s team has already collected much of the real-world data needed to populate the simulator.
The advantage of the simulator would be that the driverless car could be subjected to much more rigorous testing and that it would be much easier to detect and precisely test borderline situations. Invalid or problematic sensor data could be sent to the control unit, slippery surfaces, fog, rain and snow could be simulated by the physics engine. Unexpected behavior by vehicles, cyclists etc. could also be simulated and tested.