Spatial Cognition and Computation 2: 51–76, 2000.
© 2001 Kluwer Academic Publishers. Printed in the Netherlands.
A qualitative-connectionist approach to Robotic
Spatial Planning: the Peg-in-Hole case study
ENRIC CERVERA and ANGEL P. DEL POBIL
Computer Science Dept., Jaume I University, Campus Riu Sec, E-12080 Castelló, Spain
Abstract. Qualitative spatial reasoning for actual robots in real-world environments must
necessarily involve perceptive knowledge, since complete a-priori information of the outer
world can never be assumed, not even at a qualitative level. In this paper a contribution is
made towards the integration of quantitative data – namely, sensor signals – into a higher
level qualitative plan. This integration includes the use of neural networks to learn how to
map complex perceptual signals into a qualitative description. Sensors are used to acquire
actual knowledge of the environment and properly identify, with the help of a connectionist
system, the real state of the system. The approach is presented in the framework of robotic
tasks involving contact for which the most informative perception comes from force/torque
sensors. Empirical simulation results are provided for the chamferless two-dimensional peg-
in-hole insertion model with friction. The advantages of learning approaches over geometric
model-based techniques are discussed: our approach is simple but robust against unpredict-
able changes of task parameters, and it exhibits a gracefully degrading behavior and on-line
adaptation to new task conditions. An enhancement to incorporate a measure of conﬁdence of
the network is also presented.
Key words: connectionist learning, ﬁne motion, qualitative spatial reasoning, robotics,
This paper addresses the problem of robotic spatial planning in manipula-
tion tasks, particularly motion in contact, from a qualitative point of view.
The ﬁeld of robotic assembly and task planning must play an important
role in the automation and ﬂexibility of manufacturing systems (Kak ed.
1990). It is shown that, in the absence of precise quantitative information,
a planner is able to perform an insertion task like the peg-in-hole problem.
In this problem, a rectangular peg must be inserted into a chamferless hole
(see Figure 1). This problem has been widely used as a test case for spatial
planning with uncertainty and is also relevant to industrial robotics.