Abstract |
The current PhD thesis addresses the formulation and implementation of a methodological
framework for robot Learning from Demonstration (LfD). The latter refers to methodologies
that develop behavioral policies from example state-to-action mappings. To this
end, we study the reciprocal interaction of perception and action, in order to teach robots
a repertoire of novel action behaviors. Based on that, we design, develop and implement
a robust imitation framework, termed IMFO (IMitation Framework by Observation), that
facilitates imitation learning and relevant applications in human-robot interaction (HRI)
tasks. IMFO can cope with the reproduction of learned (i.e. previously observed) actions,
aswell as novel ones. Mapping of human actions to the respective robotic ones is achieved
via an indeterminate depiction, termed latent space representation. The latter accomplishes
a compact, yet precise abstraction of action trajectories, effectively representing
high dimensional raw actions in a low dimensional space.
Moreover, throughout this thesis, we examine the role of time in LfD by enhancing
the aforementioned framework with the notion of learning both the spatial and temporal
characteristics of human motions. Accordingly, learned actions can be subsequently reproduced
in the context of more complex time-informed HRI scenarios. Unlike previous
LfD methods that cope only with the spatial traits of an action, the formulated scheme
effectively encompasses spatial and temporal aspects. Extensive experimentation with a
variety of real robotic platforms demonstrates the robustness and applicability of the introduced
integrated LfD scheme.
Learned actions are reproduced under the high level control of a time-informed task
planner. During the implementation of the studied scenarios, temporal and physical constraints
may impose speed adaptations in the performed actions. The employed latent
space representation readily supports such variations, giving rise to novel actions in the
temporal domain. Experimental results demonstrate the effectiveness of the proposed
enhanced imitation scheme in the implementation of HRI scenarios. Additionally, a set
of well defined evaluation metrics are introduced to assess the validity of the proposed
approach considering the temporal and spatial consistency of the reproduced behaviors.
A noteworthy extension of the above regards force-based object grasping for executing
sensitive manipulation tasks. This is also treated in the current thesis via a novel supervised
learning scheme, termed SLF (Supervised Learning for Force-based manipulation).
SLF is formulated as a three-stage process: (a) supervised trial-execution in simulation
to acquire sufficient training data; (b) training to facilitate grasp learning with suitable
robot-arm pose and lifting force; (c) grasp execution in simulation. Subsequently, following
sim-to-real transfer, operation in real environments is achieved in addition to simulated
ones, generalizing also for objects not included in the trial sessions. The proposed
learning scheme is demonstrated in object lifting tasks where the applied force varies for
different objects with similar contact friction coefficients, and likewise the grasping pose.
Experimental results on the manipulator YuMi show that the robot is able to effectively
reproduce demanding lifting and manipulation tasks after learning is accomplished.
In summary, our thesis has studied LfD and has contributed with a novel approach that
introduced latent space representations to encode the action characteristics. A framework
implementation (IMFO) of our approach allowed extensive experimentation and also conduction
of HRI scenarios. The inclusion of temporal aspects in our approach enhanced it
to cope with complex, real-life interactions. Finally, the extension of IMFO with forcebased
grasping facilitated manipulation tasks with sensitive objects.
|