Saphira
Coherence is essentially the ability of the robot to maintain global world
worlds, which Konolige and Myers argue is essential for good behavioral
performance and interacting with humans.77 Finally, communication is important
because robots have to interact with humans and, as will be seen in
Ch. 8, other robots. This introduces the problem of having common frames
of reference (anchors) for communicating.
The bulk of the architecture is concerned with planning, and uses a type
of reactive planner PRS-LITE called PRS-lite for Procedural Reasoning System-lite.77
PRS-lite is capable of taking natural language voice commands from humans
(“deliver this to John”) and then operationalizing that into navigation tasks
and perceptual recognition routines. Both planning and execution relies on
the Local Perceptual Space, the central world model. Much processing is
devoted to maintaining an accurate model of the world based on the robot’s
sensors and assigning symbolic labels to regions. Saphira also divides the
deliberation activities among software agents. This provides a high degree
of flexibility. Since software agents are independent, they don’t even have to
run on-board the robot they’re controlling. In the 1996 AAAI Mobile Robot
Competition,78 robots running Saphira actually had some of their planning
reside on a local workstation, transmitted through a radio link.62 This will be
covered in more detail in Ch. 8.
The reactive component of Saphira consists of behaviors. The behaviors
extract virtual sensor inputs from VIRTUAL SENSOR the central world model, the Local Perceptual
Space. The behavioral output is fuzzy rules, which are fused using fuzzy
logic into a velocity and steer command. Fuzzy logic turns out to be a very
natural way of fusing competing demands, and is less ad hoc than Boolean
logic rules (e.g., “if x and y but not z, then turn left”). The behaviors are managed
by the plan execution of the planned navigation tasks. The fuzzy logic
mechanism for combining behaviors produces essentially the same results
as a potential field methodology, as described by Konolige andMyers. 77 The
Local Perceptual Space can improve the quality of the robot’s overall behavior
because it can smooth out sensor errors. Although this central processing
introduces a computational penalty, the increases in processor power and
clock speeds have made the computational costs acceptable.
No comments:
Post a Comment