The stereo cameras have worked poorly here, providing a low number of matched pixels. One reason for this is that many days have been overcast providing low contrast that makes it difficult for a human to see depth let alone a robot. Another problem could be calibration.
Laser is working well except in blowing snow. Filtering the data has improved laser results in low to medium levels of blowing snow but nothing works in severe cases.
The navigation system is working well but not great. With the poor results of stereo only the laser is used. This allows the possibility for an obstacle to pass unnoticed between the laser and the robot during turns. Nomad has commenced with a long distance traverse from the main camp to Nomad valley under the control of the autonomy system.
To date Nomad has avoided people 5 times and rocks 5 times while missing obstacles (requiring the e-stop to be pressed) 3 times. It has driven 8.6km under autonomous control.
Both the stereo and laser sensors have been tested on snow, rocks, and blue ice. As well, they have been tested in the following weather conditions: sunny, blowing snow and overcast. At this point I have not noticed any difference in sensor performance due to terrain type. Weather however is a different story.
The stereo cameras have in general worked poorly here. I am not positive if this is due to bad calibration or just low contrast. I am however recording lots of images for later processing. On a sunny day each pair of cameras was able to match 1500 points on the first day of testing. This is lower than the 2000-2500 points typical at the slag heaps. Since the first day, I have not had such high numbers. For many days, we have had overcast weather. This lighting creates a condition of very low contrast in the environment, such that it is difficult for humans to see depth in the snow. During these conditions, stereo performs very badly, with at most 200 points matched. This is insufficient to use for obstacle detection. Another problem encountered with the stereo cameras is a tendancy for ice to form on the inside of the front glass plate. The plates had to be removed and the ice scraped off to get reasonable images. Hopefully this does not adversly affect the calibration.
The laser has worked well in most conditions, and is usually the only sensor running during autonomy tests. The one shortfall of the laser is during intense blowing snow. It appears that the laser can be reflected back to the unit, or away - never to return - such that false obstacles are seen. Some success has been achieved by taking three samples and only accepting the largest return. The exception being that a return of maximum distance is only accepted if all three samples agree. The principle is that the laser provides a very strong indication that nothing is present from the laser unit to the measured distance, but a weak notion that there is something at the return point. This is a variant of a Hans Moravec idea which Alex Foessel suggested. It works well for light to medium blowing snow. Unfortunately for heavy snow the chance of even one of the three readings being correct is small.
Nomad has driven autonomously on three terrain types: snow (sastruggi), blue ice and blue ice with a low density of rocks. Due to the poor performance of stereo most of these tests have been done with just the laser present.
The first tests were conducted near the Chilean camp. Nomad was commanded to drive in a square, 50m on a side, and then various people would stand in front of it to serve as obstacles. After some fine tuning of the backup algorithm, Nomad was successfully able to spot people and avoid them - doing so 5 times. The square pattern was also done using only odometry. This worked well with about 2m error after driving 100m and making a 90 deg turn.
After the initial tests around the camp, Nomad embarked on its autonomous trek to Nomad valley. It drove for a distance of 3.8km to an area with large rocks on the ice. The trek was delayed here due to its interest to autonomy, landmark based navigation and locomotion tests. The trek so far has occured with little incident, mainly because there were no obstacles on the path. The main problem has been with maintaining communications over this distance.
While testing in the rocks, it became apparant that the method used by laser to determine obstacle or not was insufficient. The laser simply computed the x,y,z position of its return relative to the robot and thresholded on the z coordinate. Since the laser looks out approximately 8m in front of the robot (to allow time for the three measurement filter, processing and stopping) the z threshold had to be set high in order to avoid stopping from gently sloping terrain. This method worked well for human obstacles, but the lower rocks were not detected. To remedy this, a line was fit to the laser data. If it is assumed that most of the laser scan hit the ground, then this line represents the ground. Next, check if any point of the laser scan deviates from this line significantly. If so, its an obstacle. This new method works well and Nomad was able to successfully detect and avoid 5 obstacles with only 3 uses of the e-stop button. It did also detect 5 obstacles which did not exist.
The main problem with using only the laser for obstacle detection is that when the robot turns, it is very possible for an obstacle to move between the laser scan and the robot without ever being hit by the laser. This is the major reason for the missed obstacles.
In total, Nomad has driven 8.6 km under control of the autonomy system. While this is a large number, most of that terrain has been very benign with no obstacles to detect or avoid.
Return to Latest Updates