This project concerns efficient generation of object representations from multiple perspectives. We are developing algorithms to generate octree representation of an object from its planar projections. We use the known representation of obstacles to plan efficient motion trajectory to move an object from one location to another.

The goal of this project is to segment an image, or an image sequence, into its constituent regions such that each region is characterized by homogeneity of a three-dimensional property. Currently, we are developing segmentation algorithms that use uniformity of three-dimensional surface texture and three-dimensional object motion as homogeneity criteria.

This project is aimed at the design of a six-legged robot that is able to traverse irregular terrain mimicking the locomotion capabilities of insects. The completed robot will have a structure similar to that of an insect. In the model, each leg has three degrees of freedom, and the orientations and separations of joints are made to parallel the anatomy of the insect's legs. The robot leg movements are powered pneumatically (using compressed air) in an attempt to achieve the strength and compliance of muscle. Various robot design parameters are being obtained through experimental studies of insect locomotion. The planned controller of the robot is based on central pattern generators thought to coordinate the leg movements in insects.

The objectives of this research are analysis, development, real-time implementation, and real-world application of a new image transform. The transform is aimed at multiscale, low-level image segmentation, i.e., extraction and representation of image structure at all geometric and photometric scales present in an image. Specifically, the transform detects contours and skeletons of image regions, and identifies the cross-scale relationships among these. The scales present are a priori unknown and must be identified automatically. Application of the transform to a range of problems is investigated.

This project is aimed at the interpretation of moving, nonrigid surfaces carrying little or limited detail, with applications to virtual space teleconferencing. The objectives include delineation of moving parts of a scene, active selection of viewpoints and data acquisition, and integration of focus, shading, and silhouette information for functionality under a range of environmental conditions.

This project concerns (1) analysis of images of dynamic scenes, (2) analysis-guided synthesis, and (3) perceptual evaluation of synthesized image sequences with emphasis on computational speed, each aimed at the 3-D motion and structure characteristics relevant to navigation. The first part is concerned with integrated analysis and estimation of 3-D motion and structure parameters from multiple image cues or attributes, including those obtained during active acquisition of image sequences and those extracted from the acquired image sequence. Image synthesis is based on the new notion that the cues that contributed the most to 3-D interpretation also would contribute the most to perceptually realistic synthesis, thus suggesting an approach to analysis-guided synthesis, compression, and visualization. The perceptual evaluation tests the efficacy of analysis-guided synthesis.

This project examines the search aspects of hierarchical motion planning. Specifically, we are developing a unified view of the search performed by the FindPath algorithm, which eliminates the redundancy exhibited by most current methods. This is due in part to an improved heuristic cost function but predominantly to an algorithm for dynamically maintaining the single-source shortest paths trees of the connectivity graph.

A major limitation of current robotic systems is their inability to cope with uncertainty. This project addresses a number of issues concerning the integration of visual and physical constraints in the synthesis and execution of error-tolerant motion strategies. Object features together with their projections onto a camera image plane are used to define visual constraint surfaces. These visual constraint surfaces are used to effect visual-compliant motion, visual-guarded motion, and motion that is constrained by the simultaneous use of both visual and physical constraints. The project addresses issues of visual servo-control, as well as how existing preimage planning techniques can be extended to exploit visual constraint surfaces.

The objective of this research is to improve the quality, reproducibility, and speed of ceramic-based multichip module manufacturing. An interdisciplinary effort will focus on materials and processing issues and on automated visual inspection of tape-cast ceramic layers. This will be addressed by an experimental program that (1) characterizes the rheological properties as a function of suspension composition and time and (2) determines the microstructural variations within the layers. Insights gained from this program will be used to derive statistical models for defect occurrence, which will then drive the inspection process. The result will be improved dimensional control, reproducibility, and automated visual inspection of tape-cast sheets.

This project concerns real-time, vision-based adaptive control of robotic manipulators. The research seeks to quantify, through theoretical and experimental investigation of adaptive robot control, various instability phenomena that may arise in real-time adaptive vision feedback as a result of external disturbances and unmodeled dynamics. Such phenomena include slow parameter drift instabilities, high gain instabilities, fast adaptation instabilities, and bursting. The ultimate goal of this research is to make adaptive control of robots highly robust and therefore practical from an engineering standpoint.

This project concerns the integration of machine learning and sensor-based control in intelligent robotic systems. The research combines techniques of explanation-based control with robust and adaptive nonlinear control, computer vision, and robot motion planning. We wish to go beyond the strict hierarchical control architectures typi- cally used in robotic systems by integrating modeling, dy- namics, and control at all levels of intelligence. Our ultimate goal is to combine analytical techniques of nonlinear dynamics and control with artificial intelligence into a single new paradigm, in which symbolic reasoning holds an equal place with differential equation based modeling and control.