Cookies   I display ads to cover the expenses. See the privacy policy for more information. You can keep or reject the ads.

Video thumbnail
Micro air vehicles capable of operating
in constrained environments without the use
of an external motion capture system
are typically limited to slow and conservative flight.
As a consequence, almost all of this research
is done with rotorcraft in the hover regime.
In the robust robotics group in CSAIL at MIT,
we've developed a fixed-wing vehicle
capable of flying at high speeds through obstacles
using only onboard sensors.
The vehicle is equipped with an inertial measurement
unit and a laser range scanner.
All the computation for state estimation and control
is done onboard using an Intel Atom
processor, similar to what is found in a commercially
available netbook.
We designed a custom airplane to carry the sensing
and computation payload while still being able to maneuver
in confined spaces.
Our platform has a 2-meter wingspan
and weighs approximately 2 kilograms.
At any given time the laser can only
see a two-dimensional picture of the environment.
Laser scans are depicted with yellow points
representing obstacles, and blue representing free space.
Even with a pre-computed map, individual 2D scans
don't contain enough information to uniquely determine
the 3D position, velocity, and orientation of the vehicle.
To overcome this difficulty, we aggregate successive scans
and combine laser information with the inertial measurement
unit to perform state estimation.
Another technical challenge is efficiently
generating trajectories for the vehicle to follow.
The complicated vehicle dynamics create
substantial computational difficulties
in determining a path to fly from point A
to point B. To overcome this difficulty,
we use an approximate dynamics model
that makes it easy to map the control inputs-- elevator,
rudder, aileron, and throttle-- to corresponding XYZ
We start by connecting a set of high-level waypoints
with line and arc segments.
We then use our approximate model
to construct dynamically feasible paths
by parameterizing an offset from this underlying trajectory.
Here we demonstrate the accuracy and reliability
of this system flying through a parking garage.
In places, the parking garage is less than 2.5 meters
from floor to ceiling, creating extremely tight tolerances
for our 2-meter vehicle.
Our algorithms allowed the vehicle
to complete a 7-minute flight through the environment
traveling at over 10 meters per second, or 22 miles per hour,
covering almost 3 miles of distance
and repeatedly coming within a few centimeters of obstacles.