Development of Long-Term Memory (LTM) for ISAC


The Long-Term Memory (LTM) represents information that is stored for considerable periods of time. Remembering someone’s name, how to side a bike, or where you visited last year are all assumed to depend on long-term memory In our robot memory structures, we store motor skills as the Procedural Memory (PM).

Figure 1. ISAC humanoid robot


Development of a procedural memory structure that is a data structure that encapsulates the behaviors and forms a basis to learn new behaviors.

Project Description

Each behavior is stored as a set of trajectories in joint angle space with an indexing structure and is stored as a PM unit [1]. This indexing structure stores the initial and final Cartesian coordinates for all arm trajectories as shown in Figure 2.

Figure 2. Structure of the LTM database

The database for each PM is labeled using action and Cartesian coordinates of a robot. An action is labeled with a description of what is accomplished by the action, thus the action name can be brought into a tight correspondence with the formally represented goal of the robot.

For example, if ISAC has the goal “reach-to XYZ”, it can reason using an automatic motion generation method based on the Spatial Temporal Isomap [2]. An action labeled “reach, XYZ” will accomplish this goal. In a sense, the goal “reach-to XYZ” spawns the intention to “reach, XYZ” which directly specifies which action to take.


1 K. Kawamura, T.E. Rogers, K.A. Hambuchen and D. Erol, "Towards a Human-Robot Symbiotic System", Journal of Robotics and Computer Integrated Manufacturing (RCIM), published by Elsevier, Ltd. 2003.

2 D. Erol, J. Park, E. Turkay, K. Kawamura, O.C. Jenkins, M.J. Mataric, "Motion Generation for Humanoid Robots with Automatically Derived Behaviors", Proceedings of the IEEE Systems Man and Cybernetics (SMC), October 6-8, 2003, Washington DC.