LPR - Experiments in Interactive Intelligence

We have a team of researchers at the University of Massachusetts, Amherst that studies motor behavior in systems that learn to cope with open environments. In Developmental Psychology, our team analyze how infants use their rapidly developing motor skills build motor policies for reaching and grasping. Infants are constantly active, constantly exploring their own capabilities, constantly matching current motor strategies and skills to new tasks, and constantly growing. One of our research goals is to understand how such complex intelligent systems can develop from native mechanisms and interaction with the environment. We are exploring how knowledge regarding infant motor learning can be transferred to intelligent robots.

How Infants Learn to Plan Actions
Infants are able to plan their grasping actions both when the next action is directed toward themselves and when they are directly externally.
McCarty, M. E., Clifton, R. K. & Collard, R. R. (1999). Problem solving in infancy: The emergence of an action plan. Developmental Psychology, 35,1091-1101.
McCarty, M. E., Clifton, R. K., & Collard, R. R. (2001). The beginnings of tool use by infants and toddlers Infancy, 2, 233-256.
Berthier, N.E., Bertenthal, B.I., Seaks, J.D., Sylvia, M., Johnson, R., & Clifton, R.K. (2001). Using object knowledge in visual tracking and reaching , Infancy, 2, 257-284.

Descriptive Categories from Sensorimotor Experience
Autonomous agents make frequent use of knowledge in the form of categories -- categories of objects, activities, tasks, situations, and so on. We are developing methods that allow robots to learn categories for themselves through interaction with the environment.
Coelho, J., Piater, J., Grupen.R., Developing Haptic and Visual Categories for Reaching and Grasping with a Humanoid Robot.
Piater, J., Learning Visual Features to Recommend Grasp Configurations
Coelho, J, Grupen, R., Learning in Non-stationary Conditions: A Control Theoretic Approach
King, G., Oates, T., The Importance of Being Discrete: Learning Classes of Actions and Outcomes through Interaction
Morrison, C., Oates, T., King, G., Grounding the Unobservable in the Observable: The Role and Representation of Hidden State in Concept Formation and Refinement
Oates, T., Eylar-Walker, Z., Cohen, P., Toward Natural Language Interfaces for Robotic Agents: Grounding Linguistic Meaning in Sensors

Sensory Guidance of Grasping
We have cpnstructed an adaptive grasp controller that consists of a family of artificial potential in wrench space parameterized by contact resources and a switching function. We have shown that equilibria in one such controller correspond to optimal contact configurations for 2 and 3 contacts on regular convex polygons. Morover, we have demonstrated that a more general policy can be learned by estimating state from histories of observable variables. The haptic context of a grasp is determined by matching the dynamic behavior of the grasp controller to pre-established categories. This information reveals the geometric class of the object and can be used to address control compensation, object recognition, and contact allocation.
Grupen, R. and Coelho, C., Structure and Growth: A Model of Development for Grasping with Robot Hands.
Some Movies:

Utah/MIT hand rotating a cylinder (MPG)
Utah/MIT hand screwing-in a lightbulb (MPG)
LPR Humanoid grasp sequence (AVI)
P50 hand-arm system performing a pick-and-place operation (MPG)
P50 executing a forward grasp (MPG)

Attentional Control and Exploration We are developing a robust means of verging a stereo head that utilizes visual detectors that respond to motion, edges, and textures in a task-dependent sequence. The robot system samples these features from the visual signal to ultimately predict behavioral utility in a variety of visuomotor tasks.
Piater, J., Grupen, R., Feature Learning for Recognition With Bayesian Networks
Piater, J., Grupen, R., Constructive Feature Learning and the Development of Visual Expertise
Piater JH, Grupen RA, Distinctive Features Should Be Learned
Some Movies:

Brother, can you spare a dime? (AVI)
Visual tracking (AVI)

Exploiting Dynamics Humans exploit dynamics---gravity, momentum, elasticity, and so on---as a regular principle of motor coordination, yet most robot control techniques cancel dynamics and solve a movement task in a brute-force way. We pose a motor learning problem from which skillful movement emerges as result of interaction with the environment.
M.T. Rosenstein and A.G. Barto, Robot Weightlifting by Direct Policy Search [pdf] [ps.gz]
Rosenstein, M., Grupen, R., Velocity-Dependent Dynamic Manipulability.
Some Movies:

Robot weight lifting (MOV)

Sponsors that Collaborated on this Project

Agency	Project
NSF Research Infrastructure	Sensorimotor Development in Humans and Machines, R. A. Grupen, A. G. Barto, P. R. Cohen, N. E. Berthier, R. K. Clifton, and C. R. Beal
NSF Research Experiences for Undergraduates	R. A. Grupen, A. H. Fagg
NSF LIS	Development of reaching in human infants, N. E. Berthier
NASA Graduate Student Researchers Program	Learning To Exploit Dynamics For Robot Manipulators, Graduate Fellowship to M. T. Rosenstein
DARPA, MARS	A software control framework for learning coordinated, multi-robot strategies in open environments, R. A. Grupen
DARPA, SDR	Software, programming, and run-time coordination for distributed robotics, R. A. Grupen
NICHD	Reaching and cognition in infants, R. K. Clifton
NIMH	The development of perceptual-motor competence in human infants, R. K. Clifton

How Infants Learn to Plan Actions Infants are able to plan their grasping actions both when the next action is directed toward themselves and when they are directly externally. McCarty, M. E., Clifton, R. K. & Collard, R. R. (1999). Problem solving in infancy: The emergence of an action plan. Developmental Psychology, 35,1091-1101. McCarty, M. E., Clifton, R. K., & Collard, R. R. (2001). The beginnings of tool use by infants and toddlers Infancy, 2, 233-256. Berthier, N.E., Bertenthal, B.I., Seaks, J.D., Sylvia, M., Johnson, R., & Clifton, R.K. (2001). Using object knowledge in visual tracking and reaching , Infancy, 2, 257-284.
Descriptive Categories from Sensorimotor Experience Autonomous agents make frequent use of knowledge in the form of categories -- categories of objects, activities, tasks, situations, and so on. We are developing methods that allow robots to learn categories for themselves through interaction with the environment. Coelho, J., Piater, J., Grupen.R., Developing Haptic and Visual Categories for Reaching and Grasping with a Humanoid Robot. Piater, J., Learning Visual Features to Recommend Grasp Configurations Coelho, J, Grupen, R., Learning in Non-stationary Conditions: A Control Theoretic Approach King, G., Oates, T., The Importance of Being Discrete: Learning Classes of Actions and Outcomes through Interaction Morrison, C., Oates, T., King, G., Grounding the Unobservable in the Observable: The Role and Representation of Hidden State in Concept Formation and Refinement Oates, T., Eylar-Walker, Z., Cohen, P., Toward Natural Language Interfaces for Robotic Agents: Grounding Linguistic Meaning in Sensors
Sensory Guidance of Grasping We have cpnstructed an adaptive grasp controller that consists of a family of artificial potential in wrench space parameterized by contact resources and a switching function. We have shown that equilibria in one such controller correspond to optimal contact configurations for 2 and 3 contacts on regular convex polygons. Morover, we have demonstrated that a more general policy can be learned by estimating state from histories of observable variables. The haptic context of a grasp is determined by matching the dynamic behavior of the grasp controller to pre-established categories. This information reveals the geometric class of the object and can be used to address control compensation, object recognition, and contact allocation. Grupen, R. and Coelho, C., Structure and Growth: A Model of Development for Grasping with Robot Hands. Some Movies: Utah/MIT hand rotating a cylinder (MPG) Utah/MIT hand screwing-in a lightbulb (MPG) LPR Humanoid grasp sequence (AVI) P50 hand-arm system performing a pick-and-place operation (MPG) P50 executing a forward grasp (MPG)
Attentional Control and Exploration We are developing a robust means of verging a stereo head that utilizes visual detectors that respond to motion, edges, and textures in a task-dependent sequence. The robot system samples these features from the visual signal to ultimately predict behavioral utility in a variety of visuomotor tasks. Piater, J., Grupen, R., Feature Learning for Recognition With Bayesian Networks Piater, J., Grupen, R., Constructive Feature Learning and the Development of Visual Expertise Piater JH, Grupen RA, Distinctive Features Should Be Learned Some Movies: Brother, can you spare a dime? (AVI) Visual tracking (AVI)
Exploiting Dynamics Humans exploit dynamics---gravity, momentum, elasticity, and so on---as a regular principle of motor coordination, yet most robot control techniques cancel dynamics and solve a movement task in a brute-force way. We pose a motor learning problem from which skillful movement emerges as result of interaction with the environment. M.T. Rosenstein and A.G. Barto, Robot Weightlifting by Direct Policy Search [pdf] [ps.gz] Rosenstein, M., Grupen, R., Velocity-Dependent Dynamic Manipulability. Some Movies: Robot weight lifting (MOV)