Volume 2, Issue 1

This paper addresses the problem of segmenting perception in physical robots into meaningful events along time. In structured environments this problem can be approached using domain-specific techniques, but in the general case, as when facing unknown environments, this becomes a non-trivial problem. We propose a dynamical systems approach to this problem, consisting of simultaneously learning a model of the robot's interaction with the environment (robot and world seen as a single, coupled dynamical system), and deriving predictions about its short-term evolution. Event boundaries are detected once synchronization is lost, according to a simple statistical test. An experimental proof of concept of the proposed framework is presented, simulating a simple active perception task of a robot following a ball. The results reported here corroborate the approach, in the sense that the event boundaries are correctly detected.

We address the issue of learning and representing object grasp affordance models. We model grasp affordances with continuous probability density functions ( grasp densities ) which link object-relative grasp poses to their success probability. The underlying function representation is nonparametric and relies on kernel density estimation to provide a continuous model. Grasp densities are learned and refined from exploration, by letting a robot “play” with an object in a sequence of grasp-and-drop actions: the robot uses visual cues to generate a set of grasp hypotheses, which it then executes and records their outcomes. When a satisfactory amount of grasp data is available, an importance-sampling algorithm turns it into a grasp density. We evaluate our method in a largely autonomous learning experiment, run on three objects with distinct shapes. The experiment shows how learning increases success rates. It also measures the success rate of grasps chosen to maximize the probability of success, given reaching constraints.

The ability to learn and adapt to changing environmental conditions, as well as develop perceptive models based on stimulus-response data, provides expert human drivers with significant advantages. When it comes to bandwidth, accuracy, and repeatability, automatic control systems have clear advantages over humans; however, most high performance control systems lack many of the unique abilities of a human expert. This paper documents our first step toward the development of a novel automatic traction control algorithm using an anthropomimetic approach. The primary objective of this approach was to synthesize a high performance longitudinal traction control system by incorporating desirable human behavior distilled from human-in-the-loop (HIL) testing on a 6-DOF driving simulator. The proposed control algorithm was developed in a general framework, and applied to the specific task of longitudinal traction control. Simulation results confirm that the proposed anthropomimetic traction control algorithm provides improved performance relative to a well-tuned conventional PID-based traction control algorithm. Results are also compared with the HIL response data from a behavioral study.

This article proposes a network system composed of oscillators and capacitive couplers as an abstract neural circuit model for use in elucidating functional combinations of neural elements. The capacitive couplers produce changes in the interactions between the oscillators and enable various functions that the oscillators alone cannot realize. The functions added by these capacitive couplers include NOR-based logic operations, a falling edge detector, an astable multivibrator, and sequential pattern generation. The results of this study suggest that the incorporation of capacitive functions in an oscillatory network system is a beneficial means of extending the functionality of the system.

In this paper we present results from a study on the performance of humans and automatic controllers in a general remote navigation task. The remote navigation task is defined as driving a vehicle with nonholonomic kinematic constraints around obstacles toward a goal. We conducted experiments with humans and automatic controllers; in these experiments, the number and type of obstacles as well as the feedback delay was varied. Humans showed significantly more robust performance compared to that of a receding horizon controller. Using the human data, we then train a new human-like receding horizon controller which provides goal convergence when there is no uncertainty. We show that paths produced by the trained human-like controller are similar to human paths and that the trained controller improves robustness compared to the original receding horizon controller.