Robot haptics: Object recognition through dynamic exploration.

摘要

Much of the dexterous manipulation required in day-to-day life must be accomplished without visual feedback. Tasks such as getting keys from a pocket, changing an oil filter, or changing a light bulb all require the use of our exquisite touch capabilities. Yet, robots are lagging well behind in touch capability -- a typical robot can not even execute the varied tasks of a three-year-old! I believe the key to increasing the dexterity of robots is to pursue advances in touch software and sensor fusion.;This work pursues understanding of "robot haptics", or robot touch, using a simplified robot that pushes objects on a table. The primary goal of this dissertation is to enable recognition of objects haptically using active touch exploration.;A primary contribution of this work is based on a set of broad haptic recognition experiments that yields information on both sensor types and feature collection during haptic recognition using Naive Bayes classifiers. We analyze the data to produce a comparison of different sensor types (acceleration, force, and estimated force from commanded motor torques). Additionally, we show that sensor fusion of cost-effective sensors such as acceleration and commanded motor torque produces nearly the classification accuracy of a highly accurate (but expensive) force sensor. This has important ramifications for the design of future robot hands; future robot hands should include many low cost sensors of varying types instead of one or two highly accurate sensors.;Another important contribution is an implementation of haptic object recognition using dynamic exploration based on information gain. The robot dynamically chooses which exploration will provide the largest reduction in entropy of the probability distribution over known objects. The result is a system that autonomously explores an object to provide efficient and accurate haptic recognition that is significantly better than random exploration.