Training robots to perform real-world tasks presents unique challenges for AI because it requires matching programming instructions (software) with a physical body (hardware) in order to perform actions in the real world. Assembling the training data necessary for robots to achieve high performance on tasks can be difficult. Datasets need to be of high quality and capture the variety of possible solutions for typical robotics tasks, such as grasping an object. The demonstration data is typically generated using video or remote teleoperation of simulated or physical robots through web interfaces and virtual reality. However, it can be very time-consuming, cumbersome and costly for individual robotics labs to gather these datasets, which is a barrier to research.

The Robotics Lab at Stanford University has developed the RoboTurk crowdsourcing platform to address this challenge. By linking their smartphones to the platform, users can learn to remotely operate the physical robots located in the university lab. In a pilot study of the platform, the team collected data on the performance of three simple object manipulation tasks, such as building a tower from bowls of different shapes and sizes. The skills needed to perform these tasks are uniquely human in their required combination of reasoning and physical dexterity. Over a period of one week, a globally distributed taskforce of 54 volunteers provided over 2,000 demonstrations, totalling 111 hours of training on the tasks. The resulting dataset captured a diverse range of solutions for each of the robotics tasks. This dataset can be used to develop many different AI models.

The RoboTurk dataset is the largest one available for diverse human-operated robotics tasks, and it would not have been possible without distributed micro-tasking by volunteers. The lab has made the data openly available so that other robotics researchers can benefit from it. In the pilot study, the team demonstrated that the data could be used for developing reinforcement learning algorithms to help robots learn how to complete the demonstration tasks. This is just one of many recent data-sharing initiatives among the robotics community to accelerate progress towards more intelligent robots.

Similar initiatives:
RoboNet
DAML (domain-adaptive meta-learning)