MIT News recently posted an article “An Easier Way to Teach Robots New Skills” that features a new technique for more efficient robotic programming. The research is being done at the Massachusetts Institute of Technology and enables robots to learn new “pick-and-place” tasks and recognize unfamiliar objects with only a handful of human demonstrations.

As it now stands, robots are only trained to handle narrow tasks and have to be reprogrammed for every slight deviation. Retraining can involve hours to weeks of human labor. For example, if a robot is placed in a warehouse and assigned the task of moving cups from the shelf to the a box for shipping, if  a cup falls over, is stored upside down, or is a slightly different size, the robot will not longer recognize the object and be unable to complete its task. To remedy the situation thousands of cups of various sizes and positions must be hand labeled to reprogram the robot.

MIT researchers have developed a new training method that would allow the robot to be retrained in as little as 10 to 15 minutes. The technique involves reconstructing the shapes of 3D objects using neural network. The new neural network model, Neural Descriptor Field (NDF), teaches robots the 3D geometry of a class of items and computes the geometric representation using a 3D point cloud, which is a set of data points or coordinates in three dimensions. This allows the robot to still recognize objects with slight variants from the objects they were trained with.

Coming up with training methods that allow robots to work through variations and disparities is a vital development for robots in the workplace. Recently at the AAAS 2022 Annual Meeting the CCC sponsored a session “Robotics: Enabling not Replacing People.” One of the panel experts, Julie Shah talked about a method with a similar goal of cutting down on the hands on training robots currently require. You can learn more about her research in her slide deck posted on the CCC webpage.