Learning of Sensor-Based Arm Motions while Executing High-Level Descriptions of Tasks

Authors:
Pedro Martín;José Del R. Millán
Affiliations:
Department of Computer Science, Universitat Jaume I, 12071 Castellón, Spain. martin@inf.uji.es;Joint Research Centre, European Commission, 21020 Ispra (VA), Italy. jose.millan@jrc.it
Venue:
Autonomous Robots
Year:
1999

Citing 12
Cited 0

Real-time obstacle avoidance for manipulators and mobile robots

International Journal of Robotics Research
Integrated architecture for learning, planning, and reacting based on approximating dynamic programming

Proceedings of the seventh international conference (1990) on Machine learning
Robot motion planning: a distributed representation approach

International Journal of Robotics Research
Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

Machine Learning
Learning in embedded systems

Learning in embedded systems
Hierarchical mixtures of experts and the EM algorithm

Neural Computation
Learning to act using real-time dynamic programming

Artificial Intelligence - Special volume on computational research on interaction and agency, part 1
Learning reaching strategies through reinforcement for a sensor-based manipulator

Neural Networks
Brains, Behavior and Robotics

Brains, Behavior and Robotics
Learning to Predict by the Methods of Temporal Differences

Machine Learning
Combining Reinforcement Learning and Differential Inverse Kinematics for Collision-Free Motion of Multilink Manipulators

IWANN '97 Proceedings of the International Work-Conference on Artificial and Natural Neural Networks: Biological and Artificial Computation: From Neuroscience to Technology
Rapid, safe, and incremental learning of navigation strategies

IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics

Quantified Score

Hi-index	0.00

Visualization

Abstract

Our work focuses on making an autonomous robot manipulatorlearn suitable collision-free motions from local sensory data whileexecuting high-level descriptions of tasks. The robot arm must reacha sequence of targets where it undertakes some manipulation. Therobot manipulator has a sonar sensing skin covering its links toperceive the obstacles in its surroundings. We use reinforcementlearning for that purpose, and the neural controller acquiresappropriate reaction strategies in acceptable time provided it hassome a priori knowledge. This knowledge is specified in two mainways: an appropriate codification of the signals of the neuralcontroller—inputs, outputs and reinforcement—and decompositionof the learning task. The codification facilitates the generalizationcapabilities of the network as it takes advantage of inherentsymmetries and is quite goal-independent. On the other hand,the task of reaching a certain goal position is decomposed intotwo sequential subtasks: negotiate obstacles and move togoal. Experimental results show that the controller achieves a goodperformance incrementally in a reasonable time and exhibits hightolerance to failing sensors.