Some people need a bit of help around the house. Whether they are vision-impaired, have trouble getting around, or they just forgot where they put their keys, we think we can help. AI-Bot's goal is to create a robot from off-the-shelf parts and API's in a new way. By efficiently using consumer components, we think we can deliver cutting edge results at a fraction of the cost off other robotic platforms.
Most robotics systems try and map and understand the world in real time, using powerful but expensive GPU's for object recognition, optical flow and optical SLAM, and extremely expensive 3d LIDAR scanners to map the world around them. That's great of course, but well out of the price range for hobby developers.
Our approach is to learn from the world more slowly and generate a longer term memory of the robots surroundings. With that map in place, we can quickly perform tasks without the overhead of real-time image processing, and thus make significant savings in the price and complexity of the system.
Firstly, we use the Raspberry Pi 3 system. It's a cost effective and well supported platform for IoT and embedded development. We use an widely available tank-tracked robotics platform for its low cost and extreme stability. The rise of robotic vacuums has lowered the cost of 2D rotating LIDAR units, so we use one for SLAM mapping our environment. Recent advances in online Voice Recognition and offline Voice Synthesis API's have greatly simplified developing our hands-free interface. And lastly, Deep Learning has allowed object recognition from images, with inferences running on the consumer grade Raspberry Pi 3 hardware.
To efficiently control the robot, we plan on using a 2-stage system. First, we map out the robots home environment boundaries with LIDAR SLAM. During this process, we capture overlapping images of the environment for Deep Learning object detection using the YOLOv3 network. This neural network generates bounding boxes for the objects it finds, although it takes some time to process on the raspberry Pi 3. But as the location and direction of the robot at the time the picture was taken is recorded, we can triangulate each objects location on our SLAM-generated map. Once our map has been generated, our robot can quickly navigate using the A-Start algorithm to the location of any object in it's detection list, or let us know if it hasn't seen that object recently.
Comments