The MIT – Cornell Collision and Why it Happened
Luke Fletcher, Seth Teller, Edwin Olson
David Moore, Yoshiaki Kuwata
Jonathan How, John Leonard
Massachusetts Institute of Technology
Cambridge, MA 02139
Isaac Miller, Mark Campbell,
Dan Huttenlocher, Aaron Nathan,
Ithaca, NY 14853
Mid-way through the 2007 DARPA Urban Challenge, MIT’s autonomous Land
Rover LR3 ‘Talos’ and Team Cornell’s autonomous Chevrolet Tahoe ‘Skynet’ col-
lided in a low-speed accident, one of the first well-documented collisions between
two full-size autonomous vehicles. This collaborative study between MIT and Cor-
nell examines the root causes of the collision, which are identified in both teams’
system designs. Systems-level descriptions of both autonomous vehicles are given,
and additional detail is provided on sub-systems and algorithms implicated in the
collision. A brief summary of robot–robot interactions during the race is presented,
followed by an in-depth analysis of both robots’ behaviors leading up to and dur-
ing the Skynet–Talos collision. Data logs from the vehicles are used to show the
gulf between autonomous and human-driven vehicle behavior at low speeds and
close proximities. Contributing factors are shown to be: (1) difficulties in sensor
data association leading to phantom obstacles and an inability to detect slow mov-
ing vehicles, (2) failure to anticipate vehicle intent, and (3) an over emphasis on
lane constraints versus vehicle proximity in motion planning. Eye contact between
human road users is a crucial communications channel for slow-moving close en-
counters between vehicles. Inter-vehicle communication may play a similar role for
autonomous vehicles; however, there are availability and denial-of-service issues to
On November 3rd, 2007, the Defense Advanced Research Projects Agency (DARPA) Urban
Challenge Event (UCE) was held in Victorville, California. This event set loose, simultaneously,
for the first time, 11 full-size autonomous vehicles on a closed course. The aim of the contest was
to test the vehicles’ ability to drive between checkpoints while obeying the California traffic code.
This required exhibiting behaviors including lane keeping, intersection precedence, queuing,
parking, merging and passing.
Figure 1: The collision. (left): Skynet. (right): Talos.
On the whole, the robots drove predictably and safely through the urban road network. None of the
robots stressed the (understandably) conservative safety measures taken by DARPA. There were,
however, a number of low-speed incidents during the challenge. This paper takes an in-depth look
at one incident, the collision between Team Cornell’s vehicle ‘Skynet’ and MIT’s ‘Talos’. This
paper scrutinizes why the collision occurred and attempts to draw some lessons applicable to the
future development of autonomous vehicles.
The UCE was held on a closed course within the decommissioned George Air-force base. The
course was predominantly the street network of the residential zone of the former base with several
graded dirt roads added for the competition. The contest was cast as a race against time to complete
3 missions. The missions were different for each team but were designed to require each team to
drive 60 miles to finish the race. Penalties for erroneous or dangerous behavior were converted into
time penalties. DARPA provided all teams with a single Route Network Definition File (RNDF)
24 hours before the race. The RNDF is very similar to a digital street map used by an in-car GPS
navigation system. The file defined the road positions, number of lanes, intersections, and even
parking-space locations in GPS coordinates. A plot of the route network for the race is shown
in Figure 2. On the day of the race, each team was provided with a second unique file called a
Mission Definition File (MDF). This file consisted solely of list of checkpoints within the RNDF
which the vehicle was required to cross.
To mark progress through each mission, DARPA arranged the checkpoints in the mission files
to require the autonomous vehicle to return to complete a lap of the oval shaped “Main Circuit”
(visible in bottom left corner of Figure 2) at the end of each sub-mission. Each mission was
subdivided into 6 or 7 ‘sub-missions’. At the end of each mission, the vehicles returned to the
finishing area, where the team could recover and reposition the vehicle for the next mission. Most
roads were paved with a single lane in each direction, similar to an urban road. Several roads had
two lanes of traffic in each direction, like an arterial road or highway. One road, in the southeastern
corner of the network, was a raised dirt road constructed especially for the event.
traffic supplied by human-driven Ford Tauruses. To prevent serious crashes during the competition,
all autonomous vehicles were followed by an assigned DARPA chase vehicle. The chase-vehicle
driver supervised the robot and could ‘Pause’ or, in extreme cases, ‘Disable’ the robot via radio
link. ‘Paused’ robots could then be ‘Un-Paused’ to continue a mission when safe. ‘Disabling’ a
vehicle would kill the engine, requiring the vehicle’s team to recover it.
The qualifiers and the race provided ample opportunity for damage to the robots on parked cars,
concrete barriers, DARPA traffic vehicles and buildings. The fact that the two vehicles were not
damaged, otherthanminorscrapesinthecollision, despitehoursofdrivingemphasizesthefactthat
the circumstances leading to the collision were the product of confounding assumptions across the
two vehicle architectures. The robots negotiated many similarly complex situations successfully.
This paper begins with a brief summary in Section 2 of the robot–robot interactions during the
6-hour race. Then, to aid in the collision analysis, summaries of the MIT and Cornell vehicle
software architectures are given in Sections 3 and 4 respectively. Section 5 describes the Skynet–
Talos collision in detail, before branching in Sections 7 and 6 to provide detailed accounts of the
robots’ software state during the incident. The apparent causes of the incidents are studied here to
shed light on the deeper design issues involved. In Section 8, we draw together the insights from
the software architecture analysis to summarize the common themes, the lessons learned, and the
impediments to using these robots on the real urban roads.
Figure 2: The UCE road network. Waypoints are designated as blue dots, with traversable lanes
and zone boundaries represented as blue lines. Stop lines are designated as red circles. The Skynet–
Talos collision happened entering the Main Circuit on the bottom left.
2 Chronology of Robot–Robot interactions
The following table is a list of robot–robot collisions or close calls during the UCE. The list has
been compiled from the race day web-cast and vehicle data logs. The locations of the incidents
are marked in Figure 2.
Utah and Washing-
Cornell’s Skynet passing with IVS’
XAV-250 and Ben Franklin Racing
Team’s Ben oncoming
1h30mGeorge BoulevardBen and Team UCF’s Knight RiderSection 2.2
2h00m North Nevada and
Carolineturns across Section 2.3
3h00m White ZoneCaroline and Talos collide. Section 2.4
and Texas Avenue
Avenue Talos swerves to avoid Victor Tango’s
and Main Circuit
BoulevardSkynet and Talos collide Section 5
5h20m Utah and MontanaTalos turns across BenSection 2.6
Teams with vehicles actively involved the incidents (CarOLO, IVS and Ben Franklin Racing Team)
were invited to co-author/comment on the interactions. Any comments received are included in
the descriptions that follow.
A full discussion of the Skynet– Talos collision is given Section 5.
Diagrams have been drawn describing each incident. In the drawings, a solid line shows the path
of the vehicle, and a dashed line shows the intended/future path of the vehicle. A lateral line across
the path indicates that the vehicle came to a stop in this location. DARPA vehicles are driven by
DARPA personnel in the roles of either traffic or chase vehicles.
Videos of the log visualization for incidents involving Talos can be found in Section 9).
Skynet passing with XAV-250 and Ben oncoming at Utah and Washington
The first near-miss occurred at the intersection of Utah and Washington. Knight Rider was at the
intersection. Skynet pulled up behind a traffic vehicle, which was queued behind Knight Rider’s
chase vehicle (the chase vehicle is queued behind Knight Rider). The relative positions of the
Figure 3: After queuing behind the stationary Knight Rider, Skynet passes. (a) Visualization of
XAV-250 log when approaching Skynet (Image courtesy of Team IVS).(b) Diagram of incident.
(1): Vehicle positions when XAV-250 approaches. (2): Vehicle positions when Ben approaches.
(c) XAV-250 Camera view (Image courtesy of Team IVS). (d) Skynet(26)once XAV-250(15)had