Visual Part Identification, Localization, and Manipulation with Deep Networks

During Phase I, Symbio Robotics conducted initial development of a robust, real-time, high-performance, low-cost perception engine made possible by recent advances in deep learning. This system identifies parts in 2D or 3D images and determines their six degree-of-freedom poses, producing a full description of a scene in a fraction of a second. For this technology to be practically useful in manufacturing, thresholds of accuracy and reliability must be met, and perceptual inferences must be integrated with robust manipulator actions. In Phase II, Symbio will develop our system into a practical, complete sensory picking solution. The keys to executing this plan will be integration of information across multiple sensors and multiple images, learned grasp and motion planning, and new methods of robotic data collection. Building a multi-robot testbed will enable rapid iteration and real-world image and haptic data acquisition.