next up previous contents
Next: 9 Future Work Up: 8 Conclusions Previous: 8.3 Image based View

8.4 Visual Space Task Specification

Research in uncalibrated visual servoing has been going on for more than a decade. In [Feddema, 1994] Feddema asks why visual servoing has not yet been adopted in industry. One reason is that up to now work has concentrated on the low level servoing, i.e. obtaining a single visual alignment, and has not, in general, addressed the interaction of vision and action on a higher level (eg., how to specify tasks in a visual frame, and compose different low level servoing behaviors in order to bring about some task level goal).

In this paper we have presented three main conceptual and two practical contributions towards making visual servo control and visual frames useful for solving complex tasks (Chapter 7):

  1. Task coding and planning is done in a visual spacegif, rather than in the robot motor or 3D world space.
  2. When composing tasks we switch between different types of low level servoing, controlling different DOF's, depending on the nature of the subtask being solved.
  3. Local viewing geometry models, obtained during the course of a manipulation, are used to specify parts of the task not describable directly in terms of visual alignments.
  4. The product of our research is a vision-based interface for uncalibrated tele-assistance [Pook and Ballard, 1994] and vision based robot programming.
  5. We have demonstrated this system by solving several complex real world tasks in unstructured environments.


Martin Jägersand