Aleksandar (Alex) Vakanski

 

Robot Programming by Demonstration

The ability to transfer knowledge has played a quintessential role in the advancement of our species. Several evolutionary innovations have significantly leveraged the knowledge transfer. One example is rewiring of the neuronal networks in primates’ brains to form the, so-called, mirror neuron systems, so that when we observe tasks performed by others, a section of the brain that is responsible for observation, as well as a section that is responsible for motor control are concurrently active. The mirror neuron system represents an especially important learning mechanism among toddlers and young kids stimulating them to acquire skills by imitating the actions of adults and other kids around them. However, the evolutionary processes and modifications are very slow and prodigal, and as we further developed, we tended to rely on employing our creativity in innovating novel means for transferring knowledge. By inventing writing and alphabets as language complements, we were able to record, share, and communicate knowledge at an accelerated rate. Other innovations that followed, such as the printing press, typing machine, television, personal computers, and world wide web each has revolutionized our ability to share knowledge and redefined the foundations for our current level of technological advancement.

As our tools and machines became more advanced and sophisticated, society recognized a need to transfer knowledge to our technological inventions in order to improve efficiency and productivity, or to reduce efforts or costs. For instance, in the manufacturing industry, robotic technology has emerged as a principle means in addressing the increased demand for accuracy, speed, and repeatability. Despite the continuous growth of the number of robotic applications across various domains, the lack of interfaces for quick transfer of knowledge in combination with the lack of intelligence and reasoning abilities has practically limited operations of robots to pre-programmed repetitive tasks performed in structured environments.

Robot programming by demonstration (PbD) is a promising form for transferring new skills to robots from observation of skill examples performed by a human demonstrator. Borrowed from the observational imitation learning among humans, PbD has a potential to reduce the costs for development of robotic applications in the industry by providing ways for intuitive robot programming by end-users who are experts in performing an industrial task but may not necessarily have robot programming skills. From a broader perspective, another important motivation for development of robot PbD systems is the old dream of humankind about robotic assistance in performing everyday domestic tasks. Future advancements in PbD would allow the general population to program domestic and service robots in a natural way by demonstrating the required task in front of a robot learner.

The principal steps in solving a typical PbD problem are depicted in the figure below, although the various variants of PbD systems often include additional steps.

The probabilistic representation of human motions provides a basis for encapsulating relevant information from multiple demonstrated examples of a task. Our team developed a PbD approach that employs hidden Markov model (HMM) to probabilistically encode demonstrated tasks at a trajectory level of abstraction. The learning method is designed to generate a reproduction plan for efficient generalization from multiple demonstrations, by taking into account the variability of the motions across the demonstrated set (shown below). Furthermore, we introduced another novel approach for probabilistic learning, which employs conditional random field (CRF) in the transfer of motor skills to robots. The discriminative character of CRF is used for spatio-temporal labeling of demonstrated trajectories and for extraction of essential features for task reproduction.

Furthermore, our patented approach related to the use of vision-based control for task execution performs all the steps of a PbD process in the image space of a vision camera. The tasks are represented as trajectories of salient scene features projected onto the image plane of the camera. The task planning phase is formalized as a convex optimization problem, subject to constraints imposed by the demonstrated task, the image-based controller and the robotic platform. The main advantage of the presented learning approach is the enhanced robustness to modeling and measurement errors.

Our team developed a prototype of a commercial system for visual learning of human demonstrations. The current research efforts are directed toward improving the capacity for object tracking and pose estimation, and the existing graphical user interface of the system prototype.

 

Publications

1. M. Ghahramani, A. Vakanski, and F. Janabi-Sharifi, "6D object pose estimation for robot programming by demonstration," in Proceedings of the International Symposium on Optomechatronic Technologies, Cancun, Mexico, pp. 1–6, 2018. [BibTex] [ISOT]

2. A. Vakanski, and F. Janabi-Sharifi, Image-based Trajectory Robot Programming Planning Approach, U.S. Patent 10,112,303 (filed Aug. 25, 2016, granted Oct. 30, 2018). Entered into the Canada National Phase under 043974-P0029-CA (Apr. 2016). Based on International Patent Application PCT/CA2014/051016 (Oct. 2014). [U.S. Patent]

3. A. Vakanski, and F. Janabi-Sharifi, Robot Learning by Visual Observation, John Wiley & Sons, Inc., 208 pages, ISBN-10: 1119091802, ISBN-13: 978-1119091806, 2017. [Amazon] [Wiley]

4. A. Vakanski, F. Janabi-Sharifi, and I. Mantegh, "An image-based trajectory planning approach for robust robot programming by demonstration," Robotics and Autonomous Systems, vol. 98, pp. 241–257, Dec. 2017. [Bibtex] [Elsevier Science Direct]

5. A. Vakanski, F. Janabi-Sharifi, and I. Mantegh, "Robotic learning of manipulation tasks from visual perception using a Kinect sensor," International Journal of Machine Learning and Computing, vol. 4, no. 2, pp. 163–169, Apr. 2014. [Bibtex] [IJMLC]

6. A. Vakanski, I. Mantegh, A. Irish, and F. Janabi-Sharifi, "Trajectory learning for robot programming by demonstration using Hidden Markov Model and Dynamic Time Warping," IEEE Transactions on Systems, Man, and Cybernetics - Part B: Cybernetics, vol. 42, no. 4, pp. 1039–1052, Aug. 2012. [BibTex] [IEEE Explore]

7. A. Vakanski, F. Janabi-Sharifi, and I. Mantegh, "Transferring skills to robots for tasks with cyclic motions via dynamical systems approach," in Proceedings of the International Symposium on Optomechatronic Technologies, Paris, France, pp. 1–6, 2012. [BibTex] [ISOT]

8. A. Vakanski, F. Janabi-Sharifi, I. Mantegh, and A. Irish, "Trajectory learning based on Conditional Random Fields for robot programming by demonstration," in Proceedings of the IASTED International Conference on Robotics and Applications, Cambridge, USA, pp. 401–408, 2010. [BibTex] [IASTED]