|Figure 5. shows a typical collision-free suggestion generated by the advisor. There are nine pipe-like obstacles in the environment and the robot is avoiding collisions with itself as well. In an environment of this complexity, the advisor will respond with configuration suggestions in about one second after the operator establishes the EEF locations. In this time, the advisor evaluates one thousand trial configurations (executed on a Silicon Graphics Indigo
II R4400 150Mhz). Implementing the advisor required implementing the constraint tracking and simulated annealing algorithms described above.
Conclusions - This paper discussed the implementation of a simulated annealing optimization algorithm as part of an operator-assist interface. As an example, the paper developed an assist interface for a robot with 17 DOF called the DAW. The assist interface suggests configuration options to the systemís operator. The paper develops the computer algorithms underlying the interface. The two primary algorithms are the simulated annealing optimization algorithm and a constraint-tracking algorithm that greatly accelerates the annealing. The simulated annealing algorithm uses a weighted sum of performance criteria as an analog of energy. The algorithm then generates the configuration options with minimum energy. The
constraint tracking algorithm limits the search space of the optimization algorithm to the robotís null-space (self-motion space). The paper includes a closed-form reverse position analysis for the Schilling Titan I1 geometry as part of the constraint-tracking development.
This work shows that an optimization algorithm can quickly find collision-free configurations for highly-redundant robots in complex known environments. Equally important, however, is an evaluation of the work as an assist interface. In this case, the reference for evaluation is lofty. An experienced operator would certainly consider a number of important criteria, including: joint travel limits, joint speed limits, workspace limits, and obstacles. A good assist interface should at least match this level of expertise. Ultimately, higher-level criteria addressing issues of geometry, force, compliance, and energy will refine the suggestions. A common personal computer represents enough computing power to include these criteria in an optimization strategy. The difficulty lies in getting the optimization enough information about the robot and its environment.