Neural Network Classifies Teleoperation Data
The neural network identifies phases of tasks.
NASA's Jet Propulsion Laboratory, Pasadena, Califomia
A prototype artificial neural network, implemented in software, identifies phases of telemanipulator tasks in real time by analyzing feedback signals from force sensors on the manipulator hand. This prototype is an early subsystem-level product of a continuing effort to develop an automated system that assists in training and supervising the human control operator: the system would provide symbolic feedback (e.g., warnings of impending collisions or evaluations of performance) to the operator in real time during successive executions of the same task. Such an automated supervisory system could also simplify the transition between the teleoperation and autonomous modes of a telerobotic system. The prototype artificial neural network (see Figure 1) is based partly on the concept of the time-delay neural network, which involves preprocessing of a temporal sequence of input signals through a shift register to turn it into a temporal sequence of spatially arrayed input signals. A basic time-delay neural network contains only feedforward connections and does not exhibit adequate learning accuracy because it lacks an adequate temporal representation of the evolution of a task. To obtain better representation of the evolution of a task, the network is made partially recurrent by adding some connections from output nodes to nodes called Rcontext unitsS that are located in the input layer of neurons. The context units represent the previous state of the neural network, which state, in turn, represents the task phase executed previously. The network was trained by use of a back-propagation algorithm and training data from experimental teleoperation tasks in which a remote manipulator with a hand instrumented to measure forces and torques was controlled by the human operator via a force-reflecting hand controller and remote video monitoring of th e workspace. The tasks included insertion and removal of a peg into and from a hole, insertion and extraction of electrical connectors, and attachment of hook- and-pile pads. The network was then tested by using it to segment a force signal from the peg-in-hole task into task phases. As shown in Figure 2, the network performed the segmentation in real time, albeit with some lags and some errors. On the other hand, the network also exhibited an unexpected ability to recover after misidentifying some phases and to follow tasks, the phase sequences of which differed from those of the training tasks.
More details can be found in: P. Fiorini, A. Giancaspro, S. Losito, and G. Pasquariew, RNeural Networks for the Segmentation of Teleoperation Tasks,S Presence, Vol. 2, No. 1, pp. 1-13, 1993.
Point of Contact:
Paolo Fiorini,
Antonio Gial1caspro,
Sergio Losito,
Guido Fasquariello
Jet Propulsion Laboratory
4800 Oak Grove Drive
Pasadena, CA 91109
charles_r_weisbin@jpl.nasa.gov
The neural network identifies phases of tasks.
NASA's Jet Propulsion Laboratory, Pasadena, Califomia
A prototype artificial neural network, implemented in software, identifies phases of telemanipulator tasks in real time by analyzing feedback signals from force sensors on the manipulator hand. This prototype is an early subsystem-level product of a continuing effort to develop an automated system that assists in training and supervising the human control operator: the system would provide symbolic feedback (e.g., warnings of impending collisions or evaluations of performance) to the operator in real time during successive executions of the same task. Such an automated supervisory system could also simplify the transition between the teleoperation and autonomous modes of a telerobotic system. The prototype artificial neural network (see Figure 1) is based partly on the concept of the time-delay neural network, which involves preprocessing of a temporal sequence of input signals through a shift register to turn it into a temporal sequence of spatially arrayed input signals. A basic time-delay neural network contains only feedforward connections and does not exhibit adequate learning accuracy because it lacks an adequate temporal representation of the evolution of a task. To obtain better representation of the evolution of a task, the network is made partially recurrent by adding some connections from output nodes to nodes called Rcontext unitsS that are located in the input layer of neurons. The context units represent the previous state of the neural network, which state, in turn, represents the task phase executed previously. The network was trained by use of a back-propagation algorithm and training data from experimental teleoperation tasks in which a remote manipulator with a hand instrumented to measure forces and torques was controlled by the human operator via a force-reflecting hand controller and remote video monitoring of th e workspace. The tasks included insertion and removal of a peg into and from a hole, insertion and extraction of electrical connectors, and attachment of hook- and-pile pads. The network was then tested by using it to segment a force signal from the peg-in-hole task into task phases. As shown in Figure 2, the network performed the segmentation in real time, albeit with some lags and some errors. On the other hand, the network also exhibited an unexpected ability to recover after misidentifying some phases and to follow tasks, the phase sequences of which differed from those of the training tasks.
More details can be found in: P. Fiorini, A. Giancaspro, S. Losito, and G. Pasquariew, RNeural Networks for the Segmentation of Teleoperation Tasks,S Presence, Vol. 2, No. 1, pp. 1-13, 1993.
Point of Contact:
Paolo Fiorini,
Antonio Gial1caspro,
Sergio Losito,
Guido Fasquariello
Jet Propulsion Laboratory
4800 Oak Grove Drive
Pasadena, CA 91109
charles_r_weisbin@jpl.nasa.gov