Joseph O'Sullivan

Lessons learned from Xavier

We have been running an experiment in web-based interaction with an autonomous indoor mobile robot. The robot, called Xavier, can accept commands to travel to different offices in our building, broadcasting camera images as it travels. The experiment, which was originally designed to test a new navigation algorithm, has proven very successful. This article describes the autonomous robot system, the web-based interfaces, and how they communicate with the robot. It highlights lessons learned during this experiment in web-based robotics and includes recommendations for putting future mobile robots on the web.

Xavier: experience with a layered robot architecture

Office delivery robots have to perform many tasks such as picking up and delivering mail or faxes, returning library books, and getting coffee. They have to determine the order in which to visit locations, plan paths to those locations, follow paths reliably, and avoid static and dynamic obstacles in the process. Reliability and efficiency are key issues in the design of such autonomous robot systems. They must deal reliably with noisy sensors and actuators and with incomplete knowledge of the environment. They must also act efficiently, in real time, to deal with dynamic situations. To achieve these objectives, we have developed a robot architecture that is composed of four layers: obstacle avoidance, navigation, path planning, and task planning. The layers are independent, communicating processes that are always active, processing sensory data and status information to update their decisions and actions. A version of our robot architecture has been in nearly daily use in our building since December 1995. As of January 1997, the robot has traveled more than 110 kilometers (65 miles) in service of over 2500 navigation requests that were specified using our World Wide Web interface.

Clustering learning tasks and the selective cross-task transfer of knowledge

Recently, there has been an increased interest in machine learning methods that transfer knowledge across multiple learning tasks and “learn to learn.” Such methods have repeatedly been found to outperform conventional, single-task learning algorithms when the learning tasks are appropriately related. To increase robustness of such approaches, methods are desirable that can reason about the relatedness of individual learning tasks, in order to avoid the danger arising from tasks that are unrelated and thus potentially misleading.

Layer 3 patchy recurrent excitatory connections may determine the spatial organization of sustained activity in the primate prefrontal cortex

Anatomical studies of primate prefrontal cortex show strong and spatially modulated lateral excitatory connections among pyramidal neurons in the supra-granular layers. We study the role of patchy lateral excitatory connections in generating spatially localized pulses of sustained activity. When the patchy connections drive the firing no spatially localized sustained activity is possible. When the local connectivity is strong enough to produce sustained activity, a localized standing pulses are possible. Additional punctate stimuli boost, or extinguish the standing pulse-system depending on their spacing. When the patchy lateral connections form distinct sub-systems, spatially restricted systems of multiple pulses of sustained activity are observed over a wide range of connection strength parameters.

Multitask pattern recognition for autonomous robots

We present a method called multitask pattern recognition (MTPR) that improves the accuracy and robustness of neural net based vision systems. The method trains neural nets on auxiliary recognition problems at the same time as the nets are trained on the main recognition task. The predictions the system makes for the auxiliary recognition tasks are not used, but the internal features learned for the auxiliary tasks improve performance on the main task. In addition, the auxiliary tasks allow us to focus the attention of learning towards image features it would otherwise ignore, thereby increasing the robustness of the learned models. We demonstrate MTPR on three problems (one synthetic, two real). On these problems MTPR improves performance 10%-30%. MTPR is applicable to many pattern recognition problems.