One barrier to the incorporation of machine learning approaches in robotics is the lack of hardware platforms with the computational capacity and ease-of-use to facilitate testing new algorithms in standalone robotic devices. To address this problem we have developed bStem, a small, low-power mobile, computational platform running Linux. The board is powered by a state-of-the-art low-power Snapdragon processor with a GPU, FPGA and DSP available to accelerate algorithms. The platform also provides wireless connectivity and several adapter boards for controlling actuators, sensors and other robotic peripherals.
In addition to the hardware platform, the bStem will ship with a complete SDK to facilitate rapid development and testing of algorithms in a standalone, embedded system. The bSTEM runs a full desktop version of Ubuntu including a full-featured development environment. bSTEM uses the Python to glue together the many subsystems on board, from reading camera images and sensors to sending motor commands and plotting data into one easy-to-use framework. It also comes with optimized numerical libraries and will run a full machine learning toolkit we are developing (with Python bindings).
We will showcasing the potential of the bStem on a robotic platform. We will demonstrate learning algorithms which allow our robots to learn from a human teacher. Attendees will have the opportunity to try training the robots with complex behaviours and responses. Additionally, we will provide test machines where attendees can implement their own algorithms or tweaks using the bSTEM SDKs.
Jonathan J Hunt (Brain Corporation)
Peter O'Connor (Brain Corpporation)
More from the Same Authors
2016 Poster: Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes »
Jack Rae · Jonathan J Hunt · Ivo Danihelka · Tim Harley · Andrew Senior · Gregory Wayne · Alex Graves · Timothy Lillicrap