micropsi industries uses reinforcement learning to generate dynamic, sensor-driven behavior for industrial robots.
The micropsi industries runtime connects to robots and controls them in real time, from policies implemented in neural networks. This allows for fine-grained, dynamic control, i.e. robots reacting quickly to changing environments, fusing all available sensory information. Behaviors are trained using the best available machine learning methods for robotics, keeping training time to a minimum. Our "D3" demo video explains some of the more basic concepts, Wired also reports some details.
The micropsi industries behavior development kit, i.e. the runtime, the graphical editor, connection plug-ins for various simulation environments and "hello world" type examples, using the MNIST dataset and the OpenAI gym environment, are available for download.
First behaviors will be commercially available in the second half of 2018.