Safety Efforts for Self-Driving Trucks Should Emphasize Prevention Over Reaction

By John Marinaro, Director of Operation Safety & Testing

When people ask me about the self-driving trucks we are developing at Torc with Daimler Trucks North America, they generally ask one of three questions: 1) Are they safe? 2) How can you predict the unpredictable? and 3) How do you know they are safe? These are good questions, especially considering that Torc and Daimler have both said we won’t commercialize a Level 4 self-driving truck until we’re certain it is safe. What most people miss, however, is that safety practices for self-driving trucks — and for human drivers — should emphasize avoiding problems or incidents more than reacting to them.

Let me explain by first stating the obvious, which is that self-driving systems are not perfect. As a former NASA safety engineering executive, I understand that we cannot completely eliminate risk. In safety, our job is to understand and mitigate risk as much as possible. There are issues and parameters that we know we must understand. There are also issues we don’t even realize we don’t know, also called unknown unknowns. This is certainly the case for self-driving vehicles.

Mitigating the Risk of the Unknown

At Torc, we are building an exemplary virtual driver. Like the best human drivers, our driver can make incredibly good decisions quickly in a crisis. Exemplary drivers – both human and virtual are critical to on-road safety. About 77% of all traffic accidents are caused by driver error. These accidents are when drivers, for many reasons, are not following established safety protocols and best practices. Our virtual driver is designed with these best practices in mind – to react predictably and consistently.

Are self-driving trucks safe? Autonomous driving systems do have some technical advantages to human drivers. As mentioned, autonomous driving systems don’t get tired or distracted. Wide sensor coverage on the vehicle provides continuous and comprehensive detection of moving and stationary elements at all times – so that’s a huge reduction in blind spots that human drivers are subject to. Torc engineers are factoring in situations that are leading indicators of accidents along with specific data about high-risk areas.

We also build in detailed strategies to implement in advance of any developing risk situations. Many of these same strategies are taught to professional CDL drivers in defensive driving training programs. For example, the Take 5 System from the Smith System Driver Improvement Institute teaches drivers to aim high in steering, get the big picture, keep your eyes moving, leave yourself an out and to make sure they see you. All of this is geared toward accident avoidance, and these guidelines are industry-accepted best practices for tractor trailer drivers. Self-driving trucks can do all of this and much more.

I would like to focus on the fifth guideline from Smith for a moment: leave yourself an out. If you don’t have an escape path, then you’ve got nowhere to go if an unexpected situation happens on the road. We are also integrating the perspective that our trucks constantly face drivers who do not appreciate what it takes to drive a 40-ton tractor trailer down the road at relatively high speeds. There are physical limitations to maneuvering a large truck — such as stopping speed and turning radius — that aren’t present in passenger cars and an exit path should take these factors into account.

So how do we manage risk for unpredictable drivers on the road? How do we predict the unpredictable? We can’t control those around us, but we can accommodate them. Fortunately, we have data from studies on what increases the probability of an accident. Avoiding unnecessary lane changes, for example, will reduce chances of an accident by 12%. Of course, sometimes it’s safer to change lanes when drivers face lane closures and left exits, but avoiding accidents and close calls is a matter of best practices supported by safety data. At Torc, safety data like this allow us to program our software with information that would take a human operator decade of experience to know and master. For example, we can inform the autonomous truck to “pre-position” in a lane when a traffic situation shows pending trouble, avoiding the need for a fast change.

Some people are concerned about software and hardware failures in self-driving trucks, and at some point there likely will be failures. We are building in redundancies and fault tolerance in our design, to help manage what we cannot prevent and plan for failing safe. For this, we also take a page from NASA and aeronautics and have a “hot spare” at the ready to take over a degraded or failing system for example, adding redundancies to safety-critical systems.

Defining the Safety Envelope

Let’s also consider the driving safety envelope. Torc is defining specific driving behaviors for different traffic and environmental conditions that keep the vehicle operating within a high level of safety. This is critical to on-road safety because data show that eliminating just one factor of a developing risk event can reduce the probability of an accident by 95% or more. We classify the safety zones by color–green for safe, yellow for less safe, and orange for risky and red for dangerous zones. When we are taught to drive an automobile, for example we are told to allow a 3-second distance from the vehicle in front of us. For trucks, it’s 5-7 seconds, because it takes them longer to stop. Following the above distance guidelines would keep drivers in the green — where we want to live — at least 90% of the time. But this is not how people generally drive. A quick trip around D.C., for example, would reveal that people spend most of their time in the red zone. Consequently, there are accidents every day. Staying in the green zone would reduce accidents by more than 75%. Operating in the red zone is unsafe. You have no margin and a high chance of being forced to execute a hero response to avoid an accident.

Autonomous trucks hold the safety envelope advantage, both due to the large number of simultaneously reporting sensors and the ability to precisely measure location and act accordingly. Sensors allow the system in a self-driving truck to detect exactly where everyone else on the road is at once. It doesn’t need to take the time to look around in different mirrors sequentially, as human operators must. Moreover, the self-driving truck does not adopt human behaviors of impatience, following the crowd, or developing aggressive emotions if there is trouble on the road.

Changing lanes is a high-risk event for trucks as well as cars, and an industry guideline for changing lanes is called Take 10, which means to signal for 3 seconds and then take 7 seconds to complete the lane change. Add to this speed reduction guidelines, such as reducing speed by 5 mph for dense traffic, curves or weather conditions, and you already have an impressive package of defensive driving guidelines that greatly reduce margins for errors and mitigate risks of accidents and close calls.

We use these guidelines and others to define the behaviors of our self-driving trucks, with the target of operating in the green zone of the safety envelope as much as possible. Our goal at Torc is to lead the way in operating trucks safer, more effectively and more efficiently than most human drivers, which we hope will reduce risk and save lives. This is all possible because the driving behaviors we are talking about can be implemented consistently with the right programming.

Just as human drivers learn from experience, our self-driving trucks will gain experience. However, people gain experience one person at a time, whereas our self-driving trucks benefit from collective learning. Those sensors and computes that help the trucks perceive the situation around them can also store the data, or record the data, in case of an incident or if an extreme response is needed. We analyze all such occurrences, determine the appropriate response and update the software across the fleet. We also plan for where there is an unknown, for example, programming trucks to safely come to a stop in the case of a situation the system is not prepared to handle.

Developing the future of safety on the road

How do I know self-driving trucks are safe? Because we are building the most experienced, exemplary driver possible, supported by best-in-class hardware and computes, advanced technology, experienced truck drivers, road safety data, and industry-standard best operating practices. I imagine that before we are done, our knowledge can help teach human drivers how to be safer on the road.

Join our team

Torc’s Safety and Mission Assurance team is using a data-driven approach to developing self-driving semi-trucks. If you’d like to support our mission of helping to make the roads safer, apply today to join our winning teams.