The researchers test their algorithm using a miniature autonomous vehicle traveling along a track that partially overlaps with a second track for a human-controlled vehicle, observing incidences of collision and collision avoidance. Photo: Melanie Gonick |
Since
2000, there have been 110 million car accidents in the United States,
more than 443,000 of which have been fatal. These statistics make traffic
accidents one of the leading causes of death in this country, as well as
worldwide.
Engineers
have developed myriad safety systems aimed at preventing collisions: automated
cruise control, a radar- or laser-based sensor system that slows a car when
approaching another vehicle; blind-spot warning systems, which use lights or
beeps to alert the driver to the presence of a vehicle he or she can’t see; and
traction control and stability assist, which automatically apply the brakes if
they detect skidding or a loss of steering control.
Still,
more progress must be made to achieve the long-term goal of “intelligent
transportation”: cars that can “see” and communicate with other vehicles on the
road, making them able to prevent crashes virtually 100% of the time.
Of
course, any intelligent transportation system (ITS), even one that becomes a
mainstream addition to new cars, will have to contend with human-operated
vehicles as long as older cars remain on the road—that is, for the foreseeable
future. To this end, MIT mechanical engineers are working on a new ITS algorithm
that takes into account models of human driving behavior to warn drivers of
potential collisions, and ultimately takes control of the vehicle to prevent a
crash.
The
theory
behind the algorithm and some experimental results will be published in IEEE
Robotics and Automation Magazine. The paper is co-authored by Rajeev Verma,
who was a visiting PhD student at MIT this academic year, and Domitilla Del
Vecchio, assistant professor of mechanical engineering and W. M. Keck Career
Development Assistant Professor in Biomedical Engineering.
Avoiding the car that cried wolf
According to Del Vecchio, a common challenge for ITS developers is designing a
system that is safe without being overly conservative. It’s tempting to treat
every vehicle on the road as an “agent that’s playing against you,” she says,
and construct hypersensitive systems that consistently react to worst-case
scenarios. But with this approach, Del Vecchio says, “you get a system that
gives you warnings even when you don’t feel them as necessary. Then you would
say, ‘Oh, this warning system doesn’t work,’ and you would neglect it all the
time.”
That’s
where predicting human behavior comes in. Many other researchers have worked on
modeling patterns of human driving. Following their lead, Del Vecchio and Verma
reasoned that driving actions fall into two main modes: braking and
accelerating. Depending on which mode a driver is in at a given moment, there
is a finite set of possible places the car could be in the future, whether a
tenth of a second later or a full 10 seconds later. This set of possible
positions, combined with predictive models of human behavior—when and where
drivers slow down or speed up around an intersection, for example—all went into
building the new algorithm.
The
result is a program that is able to compute, for any two vehicles on the road
nearing an intersection, a “capture set,” or a defined area in which two
vehicles are in danger of colliding. The ITS-equipped car then engages in a
sort of game-theoretic decision, in which it uses information from its onboard
sensors as well as roadside and traffic-light sensors to try to predict what
the other car will do, reacting accordingly to prevent a crash.
When
both cars are ITS-equipped, the “game” becomes a cooperative one, with both
cars communicating their positions and working together to avoid a collision.
Steering clear of the ‘bad set’
Del Vecchio and Verma tested their algorithm with a laboratory setup involving
two miniature vehicles on overlapping circular tracks: one autonomous and one
controlled by a human driver. Eight volunteers participated, to account for
differences in individual driving styles. Out of 100 trials, there were 97
instances of collision avoidance. The vehicles entered the capture set three
times; one of these times resulted in a collision.
In
the three “failed” trials, Del Vecchio says the trouble was largely due to
delays in communication between ITS vehicles and the workstation, which
represents the roadside infrastructure that captures and transmits information
about non-ITS-equipped cars. In these cases, one vehicle may be making
decisions based on information about the position and speed of the other vehicle
that is off by a fraction of a second. “So you may end up actually being in the
capture set while the vehicles think you are not,” Del Vecchio says.
One
way to handle this problem is to improve the communication hardware as much as
possible, but the researchers say there will virtually always be delays, so
their next step is to make the system robust to these delays—that is, to ensure
that the algorithm is conservative enough to avoid a situation in which a
communication delay could mean the difference between crashing and not
crashing.
The
researchers have already begun to test their system in full-size passenger
vehicles with human drivers. In addition to learning from these real-life
trials, future work will focus on incorporating human reaction-time data to
refine when the system must actively take control of the car and when it can
merely provide a passive warning to the driver.
Eventually,
the researchers also hope to build in sensors for weather and road conditions
and take into account car-specific manufacturing details—all of which affect
handling—to help their algorithm make even better informed decisions.