2.3 Manipulation control
moves) can be recognized, it is presented in , where the recorded trajectory
is analyzed, interpreted and mapped to a manipulator.
Considering the latter approach, the space of all feasible grasp configurations
is analyzed using genetic algorithms . Since these last are not suitable for real-
time applications, neural networks have been adopted. In this way, a neuro-genetic
architecture is employed in the sense that the genetic algorithms are used to create
a training set for the neural network. In , a human-like grasp is recognized by a
biologically plausible neural network. This last is built upon a hierarchical model
for motion detection using a view-based recognition approach which is consistent
with principles in the human cortex.
Further approaches for detecting and performing human-like grasps are pre-
sented in [1, 2], where a qualitative reasoning approach to the synthesis of dex-
terous grasps is provided. An intelligent planner has been developed in order to
perform this synthesis, advantageously adopting qualitative methods instead of
analytical or numerical models. However, only coarse solutions can be provided,
since this approach is an attempt to strike a compromise in the use of qualitative
and quantitative resources.
When human-like grasps must be achieved in unstructured environments, real-
time performances are necessary and no pre-recorded trajectories are available.
Hence, some visual sensor has to be considered to reconstruct the object to be
grasped and, potentially, manipulated. In , kinematic parameters of the hu-
man grasp, such as path and preshape, are determined by the three dimensional
geometric structure of the target object, and not by the two dimensional pro-
jected image of the same object. Moreover, human object recognition is based on
identifying coarse structures rather then specific features, as underlined in .
After the grasp of an object, a typical manipulation task requires the transfer of the
same object from a starting position to a goal configuration, avoiding collisions
with obstacles, not exceeding the limits of actuators and joints, and supplying
proper grasping forces in order to ensure the stability of the grasp during the
Since a manipulation task can be subdivided in many events (i.e., change of
contacts status, change of contacts type), different control laws may be required
for each of them. The design of the feedback control should be therefore integrated
with the design of the task planner or widely interact with it.
Basically, the control law problem is the determination of the required joint
forces and torques in order to achieve the desired planned manipulation task.
Moreover, the controller needs to be robust to deal with uncertainties arising from
system modeling, actuators inaccuracy, not modeled events, unknown parameters