To Appear: AAAI Fall Symposium on Machine Learning in Computer Vision,
Raleigh, NC, October 22-24, 1993
A Vision-Based Learning Method for Pushing Manipulation
Marcos Salganicoff Giorgio Metta Andrea Oddera
Department of Computer and
University of Pennsylvania
Philadelphia, PA, USA
Laboratory for Integrated Advanced
Robotics (LIRA - Lab)
Department of Communication,
Computer and Systems Science
University of Genoa
Via Opera Pia 11A - I16145
Abstract|We describe an unsupervised on-line method for learning of manipulative actions that allows a robot to push an object connected to it with a rotational point contact to a desired point in image-space. By observing the results of its actions on the object's orientation in imagespace, the system forms a predictive forward empirical model. This acquired model is used on-line for manipulation planning and control as it improves. Rather than explicitly inverting the forward model to achieve trajectory control, a stochastic action selection technique [Moore, 1990] is used to select the most informative and promising actions, thereby integrating active perception and learning by combining on-line improvement, task-directed exploration, and model exploitation. Simulation and experimental results of the approach are presented.
Active perception can broadly be defined as the process of information gathering, organization and interpretation by the active and purposive control of sensors, effectors and computational resources in order to carry out a task or set of tasks. Closely related to the notion of active perception is the process of learning, since it holds the promise of being a general purpose method for acquiring task-specific perceptual and effector control strategies in an unsupervised way. While in principle, learning and active perception are well suited for each other, active perceptual tasks demand several properties from learning algorithms; in particular, it is desirable that the algorithms meet the following requirements:
1. That they be On-line, meaning that the system improve continually while the task is being attempted, rather than in a batched fashion which requires a separate learning and execution phase. Traditionally, many inductive learning have been batch methods, rather than on-line.
2. Closely related to the on-line requirement is the continually adaptive requirement, meaning that the system should adapt to changing task dynamics. Many inductive learning techniques are one-pass adaptive, meaning that they are allowed to adapt during the explicit learning phase described in item 1 above, but once the learning phase is over, they do not adapt to any subsequent changes in the task.
3. They should converge rapidly. Since an agent or organism has a finite lifetime, and each experimental interaction has cost in terms of time, energy and
The research was supported by ESPRIT Projects FIRST and SECOND, the Special Projects on Robotics of the Italian National Council of Research, an NSF Postdoctoral Associateship for MS (CDA-9211136), and by NSF/ESPRIT IRI-9303980
material, learning techniques should converge to a good sensing and action control policy rapidly in order to make the learning a viable alternative to hand coding of control policies.
4. They should provide an efficient exploration strategy which balances the need to explore and characterize task properties versus the need to achieve competence as rapidly as possible. Active perception systems are particularly well suited to benefit from intelligent exploration strategies since they can, by definition, determine which exemplars they create.
In this work we describe the use of a memory-based inductive learning technique which addresses these criteria and learns to perform the visuomotor task of pushing an unknown object with a single rotational point contact to an arbitrary goal point in the visual space.
II. PUSHING MANIPULATION
In many situations it is desirable to move an object from one location to another, but the object may be too large to be lifted by a single agent. Two possible solutions exist, either many agents may cooperate in lifting and moving the object [Bajcsy et al., 1991], or it may be possible for a single agent to push the object instead of lifting it. We explore the pushing case where the contact between robot and the object is single point (see Figure
1.) and the pusher remains within the friction cone of
the contact (i.e. only a rotational degree of freedom ex-
ists at the pusher/object contact point; this is enforced by notching the object at the contact point in the experiments).
Stable pushing and steering of an object to desired position in the workspace when there is only a point contact between the pusher and the object is a difficult visuomotor control problem since the relationship between the pusher and the object is unstable. Because of this, the object tends to rotate past the pusher if no corrective actions are taken. At the same time, a desired pushing direction must be achieved in order to arrive at the desired point in the robot workspace.
An additional complication is that the object motion resulting from pushing actions is a function of the frictional distribution of the object [Mason et al., 1989] on its surface of support and the mass distribution of the object, which are difficult to measure using only passive visual perception. These quantities can, in general, only be measured with active perceptual procedures [Campos
and Bajcsy, 1990; Lynch, 1993]. However, even if these