A few weeks ago, we held the 3rd iteration of the Montreal Robotics Summer School. This year the competition was fierce, and methods were very impressive. The use of ML methods in robotics has been growing, and this has been shown in the competition.
Hardware is Hard
This year, the school aimed to give more time to students to learn how to use the hardware. This is to help address a difficulty in the area of robotics and automation training and research. Robotics is hard; there are often many details to get right for an algorithm to be evaluated properly and lead to success. The performance in the final competition is likely a function of the additional time given to work with the hardware.
Growing the Community
This year, we had the opportunity to hear from many experts in robotics. The school provides a unique space for early-career robotics and ML students to learn more about the tools and people working on important research problems. This helps build collaborations and community across robotics researchers.
The Competition
After a packed few days of lectures and tutorials on robotics, control and deep learning, the students were preparing for the competition. This year the competition had three stages:
A controlled talking competition: where the goal was to get the robot to walk across a surface and get the robot over the finish line as fast as possible.
A figure 8 obstacle course with the robot controlled via, teleoperation. This task was challenging because the students needed to trains a robust policy that will overcome the obstacles in the course.
Another obstacle course, but now the robot is fully autonomous. This obstacle course is a little easier than the one in challenge 2, but the robot must use vision for localization and navigate to a goal position around the obstacle course, within 3 minutes.
The champions used a clever combination of deepRL for the low level policy and visual mapping algorthms for the high level policy to avoid obstacles and arrive very close to the goal location!
In the end, we are all winners!
The Team:
Many, Many thanks to the volunteers Florian Golemo, [email protected], [email protected] for helping with the school this year and making the competition such a success!