|My research interests cover intelligent systems in general and span robotics, haptics, computer vision, and artificial intelligence. From spatial/geometric reasoning, planning, and computation in robotics, physical simulation, and haptics, to learning (Bayesian and neural networks), and to optimization with evolutionary computation and other heuristic methods.|
From the perspective of robotics, many robotic manipulation tasks, especially automatic assembly tasks, require motion of objects in contact in order to reduce uncertainties. Automatic planning and executing such motion is necessary. We are interested in both the problem of automatic generation of contact states and the problem of contact motion planning. The two problems are intertwined. Our approach reduces the complexity by studying the topological and geometrical characteristics of contact states in the physical space of objects. With this approach, we focus on the following related subproblems: (1) properly represent discrete contact states and automatically generate the states, state transitions to obtain the contact state space; (2) investigate criteria and techniques for automatic planning of contact motions in terms of contact state transitions and contact motion within a contact state. We have achieved significant results regarding both problems for polyhedral objects. See related publications. The research has been funded by National Science Foundation (NSF) under grant IIS-9700412 (with Co-PI Robert Sturges).
We have also extended the research to contact states between general objects, including objects with curved surfaces, articulated objects, and certain deformable objects. This important extension has been funded by NSF under grant IIS-0328782. See related publications for results.
For robotics applications, theoretical concepts will be validated by experimental implementations and executions. We have collaborated with Prof. Robert Sturges of Virginia Tech on experimental verification of principal contacts and contact formations. More recently, we have started collaboration with Prof. Joris De Schutter and Prof. Herman Bruyninckx of KU Leuven of Belgium on execution of compliant motion plans and on-line (re)planning of such motions. Some results have been achieved. See related publications.
We are starting a collaborative project with General Motors for robotic surface assembly, extending the above work on contact states and compliant motion.
Our research is focused on haptic rendering involving complex contact states and compliant motion, not only between rigid objects but between one rigid and one deformable object. We have developed general methodologies for real-time (kHz) and realistic force/torque and graphic modeling and rendering. This research has been funded by NSF under grant IIS-0328782.
We have also investigated applications in assembly operations, with specific focus on optical fiber assembly, which involves assembly of deformable cable and micro/nano assembly of bare fiber.
See related publications.
Using different sensors to compensate for the uncertainties of one another is essential to accurate contact recognition, which in turn, is essential for task state recognition and for devising error recovery strategies as different contact situations may require different recovery motions (see replanning). A few researchers focused on using force sensing to verify contact states. I did some preliminary work on using vision sensing to verify contact states. These methods are all based on hypotheses-and-tests. Clearly, a key problem is how to obtain effectively the initial contact hypotheses in the first place. My approach is to generate the possible contact states based on the sensed location of contacting objects and the location uncertainties. See related publications.
This research brings out many interesting theoretical and practical issues, a lot of them still wait to be explored.
More recently, we have developed real-time algorithms for computing exact minimum distances and multiple contacts between general non-convex objects with curved surfaces of parametric representations. Our novel hybrid approach can achieve the update rate of 1KHz. It is well suited for high-fidelity interactive haptic applications, dynamic simulations, and contact motion planning. See relatedpublications.
There are generally three types of approaches to tackle the problem. One is to model the effect of uncertainties in the off-line planning process. In particular, Lozano-Perez, Mason and Taylor proposed the concept of preimages in configuration space to compute motion strategies that would not fail in the presence of uncertainties.
Another approach is to use task-dependent knowledge to obtain efficient strategies for specific tasks rather than focusing on generic strategies independent of tasks. Many strategies dealing with different types of insertion tasks fall into this category. The task-dependent knowledge can be learned explicitly by knowledge-based systems or implicitly via neural networks. The uncertainties are dealt with by either active control of forces and positions or passive reaction to errors.
Since I started working on a different approach since I was a Ph.D. student at the University of Michigan (Ann Arbor), called the replanning approach . The approach is to rely on on-line sensing to identify errors caused by uncertainties in a motion process and to replan the motion in real-time based on sensed information. In particular, I proposed a general replanning scheme, which consists of patch-planning to repair error situations caused by uncertainties and motion strategy planning to regulate motions in ways to reduce the effect of uncertainties. See the diagram below.
The replanning scheme works as follows. Given a nominal motion plan for a part held by a manipulator (for a part-mating task) which assumes no uncertainty, the motion strategy planner will generate a control strategy to execute such a plan. If the plan is stopped by an unplanned collision between the held part and the environment due to uncertainties, the patch-planner will be actived to generate a patch-plan which is to guide the held part back to its nominally planned course. A patch-plan is generated based on the analysis of the contact situation at the collision and designed to incorporate contact motions to reduce the effect of uncertainties. Once a patch-plan is generated, the motion strategy planner will generate a control strategy to execute the patch-plan, until either another unintended contact occurs, in which case the patch-planner will be activated again, or a goal state of the nominal plan is achieved (which is most likely also a contact state).
The replanning scheme is a compromise between purely model-based off-line planning (such as the preimage scheme) and rather "unconscious" low-level online control/reacting (such as many strategies for specialized insertion tasks). It simplifies the planning problem by decomposing planning into different levels and scopes (i.e., global and off-line nominal planning, local patch-planning based on sensing, and low-level motion strategy planning based on uncertainty models), treating the relatively independent issues of the problem separately. In doing so, it aims to achieve (1) efficiency , (2) generality , i.e., task-independence, by being highly modular and integrating both dynamic and static knowledge of relatively general nature, (3) effectiveness , i.e., guaranteeing or almost guaranteeing success, by being systematic and incorporating uncertainty models, and (4) flexibility , by being sensing-based and by being highly modular to facilitate the incorporation of different sensors and task-dependent knowledge upon specific applications.
A simulation package SimRep implements the replanner and simulates the assembly motion in the presence of uncertaities under the guidence of the replanner.
See related publications.