Next to overtaking on fast two-lane highways, changing lanes on dual carriageways is the most dangerous maneuver most drivers undertake. It is also one at which autonomous cars are notoriously bad, despite well over a decade spent working on the algorithms.
But progress is being made surely, if slowly, which has prompted at least one auto maker, BMW, to take a longer road towards autonomous driving than some of its competitors. It has decided that automatic lane changing will not be perfected anytime soon, and so should focus on getting that right before claiming is vehicles are ready even for SAE Level 3 autonomous driving – where the vehicle must be capable of taking over such safety critical operations under some conditions at least.
Unlike rivals such as Ford and Google’s Waymo, which plan to jump straight to Level 4 where the need for driver intervention can be eliminated under predefined conditions, BMW sees Level 3 as a critical stepping stone towards full autonomy. It regards Level 3 as a quantum leap beyond Level 2, where the vehicle can perform some functions autonomously, such as braking and steering, but not take over completely at all.
BMW seems correct that Level 3 is a greater step forward than had been appreciated, whether or not it makes sense to jump straight past it to Level 4. Even its Level 3 system, incorporated in a forthcoming electric cross over SUV called iNEXT, is not due for release until 2021 – and BMW is making a virtue of this more relaxed timescale.
The downside is that given many members of its partner ecosystem are also working with rivals, they may benefit from some of the work and implement the results sooner. BMW is building its system with help from Intel for chips, and camera sensor supplier Mobileye, which is now an Intel subsidiary. But Intel is also working with Waymo and others, so we could wonder how strong its Chinese walls are.
BMW has cited the challenge of lane changing as a factor in its decision to proceed slowly. It will have noted various advances in the underlying algorithms that have been made elsewhere, for example at the MIT (Massachusetts Institute of Technology) Computer Science and Artificial Intelligence Laboratory (CSAIL), which says it attempted to incorporate human behavior in the model with good results in simulation testing involving 16 virtual autonomous vehicles driving among several hundred others.
Of course, simulations can only go so far, and when used for driving are much more limited in value than their equivalents, for say aviation, because there are so many exception conditions and unknowns that cannot be all anticipated. There has to be support for aborting a lane change in the event that a situation changes suddenly, such as when a vehicle in an outside lane begins moving into the place the autonomous vehicle wants to enter. Most human drivers have encountered that situation at some time, and like humans, autonomous vehicles do not always cope well.
For that reason, autonomous vehicles have so far adopted a conservative approach to lane changing, often confining it to situations where it is unavoidable, such as bypassing a stationary vehicle or coming off a multi-lane highway onto a slip road.
One fundamental point is that autonomous vehicles must cope effectively and safely with a long transition period when they will share roads with others not having the same capabilities. It may well be that in the distant future, driving will become more train-like with highways acting almost as tracks guiding vehicles in convoys, aided by V2V signaling for emergency braking for example. But until this point comes, vehicles must be capable of acting truly autonomously on their own without compromising safety at all, even if efficiency and performance are slightly impaired. In any case, some vehicles at least will always want the ability to drive safely off road without the same assistance from the infrastructure.
The first lane changing algorithms were largely rule-based, which meant that they were fit only for ideal situations. They did not scale well to real roads with large amounts of traffic, because this required huge numbers of rules with multiple parameters to deal with all eventualities and proved manually intensive to tune effectively.
There was also the challenge of accuracy in the absence of clear lane markings or guidance, ensuring that the vehicle steered correctly into the adjacent lane with enough clearance but not too much. Progress was made around 2014, for cooperative lane changing, by a group at Turkey’s Technical University in Ankara – with two coupled algorithms that assumed vehicles already had features including Cooperative Adaptive Cruise Control (CACC) for safe vehicle following, lane keeping, basic control of lane changes and adjustment of vehicle distances.
The aim was to go further and support safe compulsory lane changing at some fixed position such as a traffic intersection. The first algorithm supported lane change by one vehicle in the shortest time possible while maintaining a safe distance from all other vehicles. The second algorithm then applied this iteratively back from the critical intersection to all vehicles behind, so that they all changed lanes safely and efficiently. This was one of the first algorithms supporting coordinated lane changing in which trajectories of multiple vehicles are computed simultaneously. This is a clear step towards autonomous driving of the future but still did not address the challenge of single vehicle lane changing, perhaps even just to overtake a slower vehicle.
Rule-based models have tended to be replaced by statistical systems that allow some degree of feedback, so that rudimentary machine-learning can be incorporated, to improve levels of safety and performance during simulation testing. The MIT algorithm attempts to go further in a development funded indirectly by Toyota through its Research Institute (TRI), to enable operation on the fly, with reduced data and support for different levels of assertiveness within predefined safety constraints. It generates virtual buffer zones around the vehicle, which can be adjusted dynamically to alter the level of conservatism, rather like some sports cars can change their driving style for varying levels of comfort and performance. The idea is then that as the vehicle changes lanes it adjusts its direction and speed, to maintain these buffer zones on all sides.
MIT claims this system overcomes two drawbacks of current statistically based systems, which is either being too complex for effective real time operation, or too simple to support varying levels of decision making so that in practice vehicles only change lanes when they have to.
Other systems rely on precomputed buffer zones, to ensure that risk of collision is minimized while avoiding overwhelming the in-car computer with more calculations than can be executed in the short time available. MIT has moved to dynamic buffer calculation by taking full account of the vehicle’s speed and direction, so that buffer zones are skewed in the direction of travel. This reduces the amount of data that has to be incorporated in the calculations while still being able to maintain the same minimum probability of collision. This probability can never be zero because it is impossible to rule out exceptional behavior by nearby vehicles.
Clearly though, this approach, while promising, requires a lot of testing in real roads before it can be approved for deployment. So perhaps BMW is right to hang around for another three years to wait for safer lane changing.