The DARPA Urban Challenge promises
to be a stage for innovation in the areas of robotics,
perception, and planning.
MIT has assembled a team of top researchers in these fields
and is determined to win the Grand Challenge.
With that in mind, MIT has chosen the Land Rover
LR3 because of its unique on- and off-road capabilities.
The vehicle will rely upon a layered sensor architecture
that will incorporate a coupled GPS/INS
system, a series of redundant LiDAR
scanners for short- to medium-range sensing,
and several telephoto and wide-field cameras that
will be used for object detection in the near and far
While capable hardware is necessary,
the unique nature of the upcoming challenge
places the greatest responsibility on software,
and winning will require a revolution in the approach
to the problems, ranging from perception to planning.
Whether you are in Baghdad or Boston,
uncertainty is implicit in all aspects of urban driving.
The driver must constantly be aware of a pedestrian darting
out onto the street or the car ahead suddenly braking,
as well as replan routes on the fly
to deal with unexpected road construction.
Uncertainty will similarly extend to all aspects
of the DARPA Grand Challenge.
MIT will take advantage of its unique research
and statistical state estimation and mapping,
planning under uncertainty, and statistical perception
in developing a software architecture that addresses
uncertainty at all levels of the perception, planning,
and control tasks.
At the high level, our architecture
will include a mission planner that manages the global plan,
a situational planner that is designed to rapidly handle
the uncertainty of the environment,
and a perceptual state estimator that
extracts salient information from the scene.
A map fragment database maintains a history
of the relevant scene structure and allows
the technical planner to be more aggressive when
This simulation shows a motion planning technique
applied to a sample RNDF.
Under nominal conditions, the planner
chooses the shortest path to the goal,
but is able to replant on the fly when an obstruction arises.
In the multi-vehicle case, our algorithms
predict the motion of other agents and replan short horizon
trajectories in order to avoid conflicting paths.
The path planning relies on a condensed description
of the environment provided by the sensors.
Similarly, our current vision research
is capable of real-time, 3D tracking of features
and shapes, which then provides rich and salient information
about the environment.
With video, we identify long-range movement
with particles that capture motion
that is spatially dense and temporally long-range.
The obstacles can then be tracked
using robust and optimized vision
methods that take into account the structure of the scene.
In this mockup, the system separates the road
shown in blue from the obstacles shown in red and extracts
in the far field a sign which, in turn, impacts the driving
Additionally, the mapping component of this system
will draw upon our current research in localization
and mapping, which has been demonstrated
in large, complex environments.
MIT, in conjunction with Olin College,
currently has the skills in hardware development
and construction necessary to implement
these unique algorithms.
Work by team members previously at the iRobot corporation
demonstrates the ability to deploy autonomous vehicles
that operate in rugged, outdoor terrain.
Common both to on-road and urban driving,
puddles are one challenging feature,
as they are not easily detected by standard laser range-finder
Driving in an urban environment will require highly precise
navigation at the local scale.
Using the NavCom system, we can accurately control the vehicle
based upon an IMU and by directly incorporating
occasional GPS data, can achieve centimeter precision.
Here, the system uses the vehicle's own motion
to recover three-dimensional information from the plane
or laser range data of the vehicle.
Some vehicle poses can cause a ground plane to appear
as an obstacle to the system.
Here, we show that the vehicle slows down in order
to ensure that the perceived obstacle
during the downward pitch is, in fact, the ground
plane before continuing at high speed.
Part of the DARPA Challenge requires ability
to park between stationary, unmodeled vehicles.
Here, we see a vehicle demonstrating a similar ability
by autonomously pulling into an orchard lane using a fusion
GPS/LiDAR wheel odometer InVision.
GPS is used to identify the approximate location
of the lane opening, after which local sensors take over
Another critical capability required for the DARPA
Challenge is sensing and tracking
of other moving vehicles.
Here, we see the ATV tracking the moving vehicles
ahead of it using only a single laser range-finder
and a color camera.
MIT has a long-standing history of building autonomous
platforms for a wide range of environments
that have successfully helped to pave the way for the maturing
field of robotics.
We are looking forward to applying
our technical experience to the 2006 DARPA Urban Challenge.