Learning Sensory Patterns and Skills from Human Demonstration

A central focus of my research is sensory perception and intelligent control, especially as it relates to the human-machine interface. The current focus has been on how humans and robots interact and how the human-to-robot programming interface can be made more natural for the human operator. The work is motivated by the vision of intelligent robot tools and assistants, which operate with a human worker, making his job easier, extending his capabilities, and allowing him to avoid hazardous environments. The robot may be used to perform the more menial or more dangerous tasks. If it is a new task not performed previously, the robot may need some quick instruction or task refinement (i.e., programming), most likely done at the work site with instruction provided by the actual user rather than a specialized robot engineer. The robot may be positioned next to its human operator or it may be located at a distance. In either case, the robot will be more functional if it can be easily reprogrammed for new jobs by the person that uses it.

The goal of my dissertation research was to find a fast and natural method of transferring human assembly skills to robots, in such a way that the skill execution can be performed successfully in semi-structured environments. The learned skill had to support movement in 6 degrees of freedom and be robust enough to withstand uncertainties in the environment (e.g., position and orientation uncertainties). The transfer method had to be fast and efficient and provide a natural, intuitive interface for the human demonstrator so that it can be used by end-users of the robot. For this reason, two key decisions were made. First, a strictly sensor-based approach was used instead of one requiring detailed geometric models. Second, a programming by demonstration approach was used. Because the skill transfer involved assembly/contact operations, force sensor signals were used as the robot's haptic sense. The skills were demonstrated telerobotically, using force reflection, through the use of a PHANToM haptic display device as a hand controller for driving the robot.

The mathematical framework for the skill model is the hybrid control model proposed by Brockett [1], in which a discrete event controller interacts with a continuous-time plant (the robot). The discrete event controller issues discrete commands to the robot, in the form of reference velocity commands. Continuous signals from the robot's force sensor are mapped into discrete states, termed single-ended contact formations (SECF), which describe qualitatively how a grasped object is touching its environment. A change in SECF triggers a new controller command, so a critical point is to learn the mapping from the force signals to the SECF states, in a fast, efficient manner. Unlike similar work (e.g., [2]), the model also allows for force control commands to be combined with reference velocity commands.

Skill acquisition involves the learning of three functions: (1) the mapping of force sensor signals to SECFs, (2) the sequence of SECFs, and (3) the transition commands which move the robot from the current SECF to the next desired SECF. The first function is acquired using supervised learning. The operator demonstrates each SECF while force data is collected, and the data is used to train a state classifier. The operator then demonstrates a skill, and the trained classifier is used to extract the sequence of SECFs and the transition commands which comprise the rest of the skill. To provide robust execution of the skill, the operator can repeat the demonstration using different (but still successful) sequences of SECFs. The skill is stored in the form of a graph, where each node represents a different SECF, and the edges represent transition commands. The robot executes the learned skill, using the graph. As the robot moves and makes contact with its environment, the current SECF is classified in real-time from force signals. If necessary, the sequence can be changed on-line, with the appropriate transition commands automatically generated.

Two key elements of this research distinguish it from similar work. The first is the pattern recognition technique used to classify the SECFs, which produces very fast but good results. Fuzzy logic is used to model and recognize patterns in the force vectors, with membership functions being generated automatically from the training sets. Reference [3] describes this process for the static case; [4] addresses the dynamic case. In addition, a neural network classifier has been developed, and the performance of the two classifiers is compared in [5]. Another unique element of this work is the interactive interface between the human teacher and the robot student. During the skill demonstration, feedback is provided to the human teacher by classifying and displaying the SECF sequence in real-time. The teacher can then assess whether the robot is learning the intended skill technique. Also, the interface includes an interactive method for the teacher to refine the learned sequence, which allows skill acquisition to take place even if the demonstrated task is flawed.

This research domain draws from work in several related areas: sensory perception, intelligent control, machine learning, fuzzy logic, neural networks, and human-machine interaction. The central theme is understanding and using sensor signals. The work continues with the following research projects:

(1) As a continuation of the assembly skill transfer, clustering is used to capture classifier training data from the demonstrated sensor signals.

(2) A virtual environment is created for transferring assembly skills to robots. Collision detection of the geometric models provides identification of the single-ended contact formation, which is used as a position-independent link between the virtual and real environments.

(3) Interactive training is applied to mobile robots. Clustering and pattern recognition provide sensor-based cues for learning a task. Potential applications include remote inspections, hazardous waste handling, and intelligent assistive devices for disabled persons.


[1] Brockett, R.W., "Hybrid models for motion control systems," in Essays on Control: Perspectives in the Theory and Applications, H.L. Trentelman and J.C. Willems, Eds., chapter 2, pp. 29-53, Birkhauser, Boston, MA, 1993.

[2] Hovland, G.E., P. Sikka, and B.J. McCarragher, "Skill acquisition from human demonstration using a hidden markov model," in Proceedings of the 1996 IEEE Intl. Conf. on Robotics and Automation, Minneapolis, MN, Apr. 1996, vol. 3, pp. 2706-2711.

[3] Skubic, M. and R.A. Volz. "Identifying Contact Formations from Sensory Patterns and Its Applicability to Robot Programming by Demonstration", in Proceedings of the 1996 IEEE/RSJ Intl. Conf. On Intelligent Robots and Systems, vol. 2, pp. 458-464, Osaka, Japan, Nov., 1996.

[4] Skubic, M. and R.A. Volz. "Learning Force Sensory Patterns and Skills from Human Demonstration", in Proceedings of the 1997 IEEE Intl. Conf. On Robotics and Automation, vol. 1, pp. 284-290, Albuquerque, NM, May, 1997.

[5] Skubic, M., S. Castrianni and R.A. Volz, "Identifying Contact Formations from Force Signals: A Comparison of Fuzzy and Neural Network Classifiers", in Proceedings of the 1997 IEEE Intl. Conf. On Neural Networks, vol. 3, pp. 1623-1628, Houston, TX, June, 1997.



January, 1998