Sainsbury trialing Dearman liquid air engine in refrigerated truck
Mitsubishi Motors forecasts 41% drop in retail sales in Japan over improper fuel consumption testing

Avoiding Obstacles Safely at High Speed: An Optimal Control Approach for Driving Heavy Autonomous Ground Vehicles Close to their Dynamic Limits

by Paul N. Blumberg, PhD

No one who follows automotive news and technology trends can be unaware of the intense research and experimentation taking place related to vehicles that drive themselves, i.e., “Autonomous Vehicles.” Among other potential benefits, these vehicles offer the promise of increased safety on the nation’s roadways and personal mobility to those who are not able or do not wish to drive themselves.

The Society of Automotive Engineers has already developed Standard J3016 which describes six different levels of “driving automation” increasing from Level 0 - No Automation, to Level 5 - Full Automation. At Level 0, a human driver performs all driving tasks, even if assisted by warning or other momentary intervention systems. At Level 5, a system comprising sensors, actuators and control algorithms performs all driving tasks such as steering, braking, acceleration and monitoring the vehicle and roadway under all modes of driving. At Level 5, a human driver is no longer necessary.

Obviously, there is a long and difficult path to progress from Level 0 to Level 6. In fact, as the level of automation goes up, the research and development task increases in a very rapid and steep fashion due to the increased sophistication of both the hardware and software that are required.

Recently, Raj Nair, Ford Motor Company’s Vice President for Product Development was quoted in Automotive News saying that “it is credible that an autonomous vehicle at SAE Level 4 [one step below full automation] will hit the market by 2020.” Even if this proves to be an optimistic projection, it attests to the level of enthusiasm that surrounds the development of this highly disruptive new technology.

Mr. Nair noted that the most significant needs were “improved sensor capability and the development of software that can read the road as a human would.” To those following the development of this exciting technology, it is clear that along the development path to higher levels of automation, many safety aids and improvements will emerge and will work their way into commercially available products.

It is self-evident that the US military, specifically the Army, would be highly interested and supportive of efforts related to the development of high speed, large and heavily loaded Autonomous Ground Vehicles (AGVs) to avoid sending humans (i.e., soldiers and/or military specialists) on dangerous missions where injury and death might be possible outcomes. In certain areas of this technology they would naturally be in a forefront position, given the difficult circumstances under which they would be operating military AGVs. Of additional interest is the fact that the principles and features of this technology could find applicability in the civilian heavy duty sector, as discussed in brief at the conclusion of the article.

As the US Army’s Center of Excellence for Modeling and Simulation of Ground Vehicles, the Automotive Research Center (ARC), led by the University of Michigan, is in a forefront position to develop innovative and advanced methods of controlling AGVs under almost all conditions.

In recently completed work, Model Predictive Control (MPC), which includes optimization of overall vehicle performance with respect to a specific mission, has been employed to demonstrate that an AGV can reach a pre-chosen target position in terrain that is laden with fixed obstacles without any collisions with these obstacles and with no loss of dynamic stability, which could lead to vehicle rollover or skid.

The researchers involved in this work have considerable experience in this field and have published earlier technical papers on their work. The effort recently completed represents the most sophisticated approach and use, to date, of advanced MPC methods for obstacle avoidance and maintenance of vehicle stability. In prior work, the vehicle speed was held at a constant value. However, in the latest work the vehicle can be controlled to accelerate and decelerate as well as change its direction of travel (steer) in conjunction with its intrinsic capabilities in an optimal manner that ensures safe travel from an initial position to a target or “goal” in the minimal possible time.

In this work, the AGV is operated and controlled in what is termed an unstructured environment. In practical terms this means that there are no lanes or other markers to follow and no traffic rules that must be obeyed. In their work, the researchers have defined the mission of the AGV as moving safely (i.e. obstacle avoidance and dynamic stability) from its initial position to a predesignated goal position as quickly as possible.

Between the starting point and the goal, there are obstacles whose location, size and shape are not known in advance. The only information as to the location and nature of the obstacles is obtained through the use of a LiDAR (Light Detection and Ranging) sensor that is mounted on the front of the vehicle, giving it a very wide frontal field of view extending to about 100 meters in distance. The LiDAR sensor illuminates its field of view with laser light, which can detect the distance to an obstacle and most, but, unfortunately, not all of the precise boundaries of the obstacle. The uncertainty introduced by this lack of 100% precise information as to the shape of the obstacle is accounted for in the control algorithm.

As for the dynamic stability of the vehicle, this is achieved by ensuring that a positive vertical force exists between all of the vehicle tires and the ground at all times, which has the effect of not allowing a single tire to lift off the surface at any time during the vehicle’s travel. This requires that a full, well-calibrated transient dynamic model of the vehicle—including mathematically based description of the powertrain, steering system and brakes, as well as both the longitudinal and lateral load transfers due to acceleration/deceleration and turning forces—be embedded within the control software.

In actually determining the optimal vehicle trajectory, the total region of potential travel, as determined by the feedback from the LiDAR sensor, is broken up into various categories consisting of a number of sub-regions. Some of the sub-regions are “open” and free of obstacles and some are boundary regions terminated by obstacles. The vehicle is propelled forward in a sequence of multi-stage time segments or “prediction horizons”, during each one of which the vehicle may be accelerated, decelerated and steered according to the prediction and instructions of the predictive control optimizer.

The control optimizer is formulated so that the vehicle will reach its destination safely (with respect to both speed and dynamic stability) in the shortest possible time. The detailed mathematical formulation is complex and will not be discussed in detail in this article. A full description can be found in the technical paper the researchers have published[1] as well as in their prior work.

The researchers conclude that Model Predictive Control allowing both variable velocity and steering control, as opposed to the constraint of constant speed, not only improves the performance of the AGV by allowing it to operate closer to its dynamic limits but also enables its safe travel through and around a set of obstacles that might not be traversable at constant speed with steering control alone.

For example, in a 500-meter trajectory specified by the researchers, the AGV with variable speed control—the speed varying between approximately 10 m/sec and 30 m/sec after the initial acceleration from a stopped position—reaches its target about 7% more quickly than a vehicle limited to a constant speed. This can be decisive in mission-critical or hostile situations. Furthermore, at an enforced constant speed of 20 m/sec, the vehicle collides with the obstacle, as shown in the figure.

Comparison of the Trajectories of Vehicles Employing a Variable Velocity Controller vs. a Constant Velocity Controller

Future work will investigate the use of improved software algorithms that explicitly take into consideration uncertainties in both the models and in sensor measurements. This is expected to improve the overall robustness of control. Applying MPC to moving obstacles is another direction for future research as well. Of course, as sensor technology improves, incorporation of their improved characteristics will also be implemented and evaluated.

In the heavy duty truck commercial sector there are also obstacles (such as the vehicle directly ahead or a vehicle that suddenly cuts in or simply the presence of potentially dangerous debris in the road) combined with an economic and logistical incentive to safely reach a specific location or goal in a minimum amount of time. Dynamic stability is also of concern as rollover or jack-knifing are to be avoided at all costs.

However, the environment for these vehicles is structured, i.e. there are generally well-defined lanes and traffic rules exist (both explicit and implicit) that are to be followed. Although fully autonomous heavy duty vehicles are not foreseeable in the very near term, it is evident that the principles of Model Predictive Control, combined with appropriate sensors and algorithms, could serve to produce significant passive aids for the truck operator in the form of information and alerts. These aids would entail both speed and steering control advisories for collision avoidance and dynamic stability. In extreme cases of immediately impending danger, they could be implemented to intervene actively in order to initiate the required extrication maneuvers.

Dr. Blumberg is currently a Visiting Research Scientist, Mechanical Engineering, College of Engineering at the University of Michigan, Ann Arbor.


  • [1] “An MPC Algorithm with Combined Speed and Steering Control for Obstacle Avoidance in Autonomous Ground Vehicles,” Liu, J., Stein, J.L., Jayakumar, P and Ersal, T., Paper No. DSCC2015–9747, Proceedings of the ASME 2015 Dynamic systems and Control conference, October, 2015


The comments to this entry are closed.