GTAR - 2002

2002 Aerial Robotics

Raw data files

BeforeCalgary

The Vehicle

Rotor Diameter 10.2 ft
Length 11.9 ft
Gross Weight 205 lb
Payload >66 lb
Engine Gasoline, 2 Cycle 2 Cylinder, Water Cooled 246cc Displacement, 21 hp
Endurance 60 min

Avionics Enclosure

In parallel with waiting for sensor items and parts, the avionics enclosure was starting to take shape around October 2001. The payload of the Rmax is around 45 kg. This allowed the design of the box to be made flexible enough to replace sensors and add sensor / computer modules to the platform. The mass of
the final avionics box was around 15 kg.

Navigation, Control and Simulation

Some of our earliest tests involved tuning the navigation system. Much of this work was performed on the back of a truck where the avionics box and all sensor antennae were mounted and driven around campus correcting and tuning the nav filter.

Nav tests with truck. This is in the YMCA parking lot behind the Aerospace building. The control systems and navigation filter were developed and tested using an in-house simulation tool called Esim.

Simulation was used extensively from testing the navigation, control to developing device drivers and use as a ground station. Some details of the simulation tool and how it was used used may be found in

  1. Simulation and Development Environment for Multiple Heterogeneous UAVsSuresh K. Kannan, Adrian A. Koller, Eric N. JohnsonAIAA Modeling and Simulation Technologies Conference and Exhibit, Providence, Rhode Island, August, 2004Keywords: simulation, software

The helicopter uses an adaptive controller based on feedback linearization and is documented in

  1. Adaptive Trajectory based Control for Autonomous HelicoptersSuresh K. Kannan, Eric N. JohnsonThis paper won the Best Student Paper of Conference award21st Digital Avionics Systems Conference, Irvine, CA, October, 2002Keywords: adaptive control, autonomous, flight test results, helicopter control

Some of the many uses of simulation and configurations are shown below.

Single Computer full simulation including model, nav, control and ground station. Usual configuration during development and testing.

This configuration is where all flight code runs on actual hardware but sensors are simulated. The flight computer thinks it is in the air.

This configuration shows testing where all sensors are turned on and the flight computer is talking to actual sensors. In the lab however, GPS is unavailable and hence only that signal is emulated and sent to the flight computer. This has the disadvantage that one may not fly maneuvers but it does test all the sensor drivers, and nav filter. Hence, the flight computer thinks it is sitting still.This configuration is the closest one comes to flight in the lab.

Image Processing & Tracking

The original goal was to use the fast Texas Instruments DSP board and daughter card for Image Processing. In fact many of the algorithms actually worked very well. However, the TI boards interface for communication was quite cumbersome. The only truly supported interface was via the Parallel Port to Windows. Although talking via parallel port to a Linux Machine was partially achieved, in part due to an open source for BSD driver. If used, the communication path to the TI board would have become to cumbersome. TI:BSD driver:Wireless ethernet packet and finally to Ground station. We hope to master the TI board interface soon. The final decision for aerial robotics, was to use an open source library called OpenCV from Intel. It provided many image processing primitives and runs on NT and Linux. Most of the image processing was developed on Windows but during the latter stages it was tested and run primarily on Linux. The secondary computer acted as the Image Processing system with a minimal installation of Redhat 7.3 on a 512MB flash disk. A video server from www.axis.com was used to take analog camera input and ftp images at 5Hz to a ramdisk on the secondary flight computer (linux).

Overview of the vision system.

The vision system consisted of two parts, the image processing and tracker. The image processing was deliberately kept simple and acted on one image at any given instant. The image processor picked the best candidate for a building (rectangle) and iarc symbol, scored them and passed their location onto
the tracker. The tracker kept a record of events happening temporally and kept track of confidences of building, window and symbol locations. At any given instant, the tracker, knowing the position and attitude of the helicopter and camera can estimate the expected size of the window and building. This information is then used by the image processor to pick rectangles that best matched the area
criteria.

Overview of the image processing test.

A sample of the onboard video (about 16 MB) The image processing was tested in the lab using playback of onboard video, whereas the tracking algorithm was tested using simulated images.

In order to test the tracker, a more controlled video sequence was generated by pointing the helicopter camera at a simulation. Since the scene generation and helicopter model was already available, a virtual camera was placed at the same location and orientation as the real camera. The resulting scene window was also customized to reflect the field of view of the real camera. A laptop
running the simulation was placed in front of the helicopter camera, and a hardware in the loop simulation was conducted to verify the performance of the tracker. The entire Level 1 and Level 2 missions were simulated in this fashion. This very same setup was run in flight. Only three flight tests primarily for vision were required

Flight Tests

The final flight configuration was essentially a small change from the hardware in the loop simulations in the lab.

Movies

Calgary Movies

Email From Calgary

Calgary Olympic Park,
Calgary,
Canada.
August 2, 2002.
Sent: Friday, August 02, 2002 3:59 PM
Subject: Results from the Aerial Robotics Competition

All,

This is a brief summary of the Georgia Tech experience at the 2002 International Aerial Robotics Competition in Calgary, Alberta. The first major event was a chance to create new case law for the import/export of UAVs – obviously not one of our goals. By sheer chance, we got a U.S. customs inspector who had just become a “specialist in UAVs” and detained our vehicle at the border. After a massive effort by Gary Wolovick, Dan Schrage, Robert Michelson, the team, and many others we received word from the State Department that we were free to go and legal (as in we had done nothing wrong, although they recommended that we seek prior approval next time) 28 hours after it was detained. To my knowledge, this is a new requirement for this class of vehicle, and it appears to be at least partially due to 9/11 that this occurred. The team caught up fast, actually getting in a few practice flights the same day the vehicle was released. The following day was “static judging”, which is used to prioritize flight times for contest day. Georgia Tech was fourth, which was good enough to allow the team to pick its favorite time slot:first, 7am. Practice flights that afternoon included flight in windy and gusty conditions, and the vehicle performed beautifully. Discovery channel was filming these flights, so look for this in the future.

Contest day: The GTMax was in the air for virtually the entire 1 hour of Georgia Tech’s allotted time (because a mission attempt can start
and end in the air). The team’s goal was to achieve the Level 2 mission, which includes locating a particular building in a group of three buildings, and then locating at least one entry point in that building. This must be done with a completely automated air vehicle and image processing system. The bad news for everyone was that there were a number of emitters at the site (radio stations and other functions) which led to degraded GPS and datalink performance. Loss of GPS signal and datalinks led to at least 6 aborted mission attempts. Three of the attempts proceeded far enough to successfully automatically locate all three buildings. Two of the attempts proceeded far enough for the vehicle to descend and circle the buildings, obtaining “close up” pictures of the windows and identifying the correct building. Just to keep things dramatic, it was on the final attempt during the allotted time that the GPS and datalinks held out long enough to map all three buildings. .and just to make it extra dramatic the low fuel light came on while mapping that last building. After several hours of looking at the data produced during the two mission attempts that included results for entry points, the following was determined:

On the first, the mission was aborted (due to GPS on that one) before the vehicle got to even try to map the correct building. On the one mission attempt that got to map all three buildings, the correct building was identified based on a specified symbol (about 1 meter in diameter) located on the building. This building was only 12 feet from another building, and the automated system unfortunately picked an entry point from this neighboring building as its solution.

In summary: The vehicle and autopilot flew great, the buildings were located automatically several times, window mapping was quite precise in position, but due to GPS and datalink dropouts only one full mission attempt was completed – and due to a stroke of bad luck it picked a window from a building 12 feet away.

Additional completed mission attempts almost certainly would have made it. Due to the interference problems mentioned above, no other team did any notable flight attempts on the contest day. Most didn’t leave the ground because they couldn’t use their GPS and/or their manual control links were not achieving enough range.

Although we didn’t get Level 2, we did achieve plenty – in fact, far more than enough to finish in first place for the second year in a row.

Thank you to all of you for supporting the team, including our
sponsors:
Lockheed Martin, NovAtel, Bahr Avionics, Texas Instruments and Guided Systems Technologies

And finally, our gratitude to Ted, Doreen, Beth Pritchett and Sparky for hosting us in
Calgary. Without Ted’s universal gadget box most of us would have become icicles.

Pictures

Here are some pictures of the team after the competition and of Alison and Mike on their return trip.

Relevant Papers

2002 Entry Paper

  1. Adaptive Flight Control for an Autonomous Unmanned HelicopterEric N. Johnson, Suresh K. KannanAIAA Guidance, Navigation, and Control Conference and Exhibit, Monterey, California, August, 2002Keywords: adaptive control, autonomous, helicopter control, neural, trajectory control
  2. Adaptive Trajectory based Control for Autonomous HelicoptersSuresh K. Kannan, Eric N. JohnsonThis paper won the Best Student Paper of Conference award21st Digital Avionics Systems Conference, Irvine, CA, October, 2002Keywords: adaptive control, autonomous, flight test results, helicopter control