Real - Time Recognition and Visual Control :

Real-time vision for the control of actions, real-time control of vision processing, learning, and high-and low-level perception continue to be central themes at Rochester. e. We continue to integrate vision into behaving robotic systems. The main developments this year have been in software: the RCCL robot control program has been successfully ported to the Solaris environment, so that it can run in parallel with visual sensing on the multi-processor Sparc-Station2000. We have acquired several more digitizers for the Sbus, which supplement our continuing use of the MV200 pipeline image processing boards Vallino 1995]. 2 Deictic Visuo-Motor Control Humans in control loops can be necessary in many situations , but they can be problematic as well. The method of deictic control is a tool for semi-autonomous systems, in which relatively low-bandwidth communication by a human is married with high-bandwidth local control and servoing to produce behavior that is more accurate, robust to delays, less demanding on the human controller than teleoperation and less demanding on the artiicial intelligence capabilities of the controlled system than full autonomy. The deictic paradigm has been explored in the context of manipulation by Pook and Ballard Pook 1995a,b; Pook and Ballard 1994; Pook and Ballard 1995]. Here it takes the form of hand gestures encoded by an exoskeletal master, and interpreted by the controller to give the intentions of the human and some relevant geometric parameters. A pointing gesture means \move in a straight line in the indicated direction", for instance. Local servoing produces the straight-line motion and the move is guarded so as to stop at any obstruction. Pook's work demonstarates that deictic control results in a much lower amount of human involvement with the control process, but it is involvement that takes place at critical junctures (at the splice points between sub-jobs, for example). Her work also demonstrates a robustness against delays (up to eight seconds in one demonstration) between commands and robot execution. Commands were issued both using direct vision and with remote views displayed on a virtual reality helmet worn by the human. Computer-naive and completely untrained subjects could control the 16 degree of freedom Utah hand and Puma arm combination to open a door using the deictic method, but an untrained operator is hopeless and dangerous using straight teleoperation. The deictic control strategy was the basis for an ARPA DURIP grant to Nelson, Brown, and Prof. Roger Gans of Mechanical Engineering, …