Robot assembly is an essential operation in manufacturing processes, and future
generations of robots are intended for use outside of industrial environments. Traditional
robot programming for assembly tasks can be time-consuming, as it is necessary
to anticipate all possible situations in advance and create a usually complex robotic
program. This requires profound programming skills and an in-depth understanding of
robotics, which limits the use of robots to industrial applications with predictable, welldefined tasks. For the broader use of robots, rapid, intuitive, and effective solutions
for transferring human knowledge to the robotic system should be developed.
In this dissertation, we examine the potential of human-robot collaboration for
incremental learning of assembly tasks, focused on exploitation of variable stiffness.
We design a robot learning system adaptable to various tasks and settings as it can
be taught in two complementary ways: from demonstrations and by following environmental
constraints. Assembly tasks can be improved over time through various
collaborative processes between a human operator and a robot: gradual refinement of
the robot control policies through kinesthetic guidance and providing additional policies
to teach the robot system how to handle unpredictable situations that can arise
during the assembly.
In the first part of the dissertation, we focus on enhancing methods for effective
kinesthetic teaching to facilitate their application in assembly tasks. The quality of the
demonstrations is crucial for accurate and correct assembly. Our work is based on the
observation that the performance of assembly tasks can be significantly improved in a
few iteration steps, much like humans improve their skills through repeated exercise
and receiving feedback. To this end, we develop a procedure for gradually refining
existing trajectories through kinesthetic guidance along virtual tunnels at arbitrary
speeds, with the ability for humans to demonstrate initial movements and refine them
until satisfactory robot execution is achieved. Since complex movements can be more
accurately demonstrated at a low speed, it is also necessary to allow for a separate
refinement of the spatial and temporal part of the task. We propose to first refine the
spatial part, i.e., the shape of the trajectory, with an arbitrary speed. The desired
speed is learned in the last stage of the procedure. In our approach, both changing of
the shape and the speed of the trajectory is implemented by moving the end-effector
back and forth inside the virtual tunnel.
In the second part of this dissertation, we address the challenge of ensuring the
robust execution of assembly tasks in unstructured environments where unforeseen
situations may cause errors even if the robot is carefully programmed and optimized.
We propose a collaborative approach to handle exception scenarios, which consists
of a multi-stage process that enables the robot to recover from errors and continue
with the assembly task. First, the system remembers the context in which the error
occurred and observes how the operator handles the situation using our incremental
policy refinement method. The robot then uses statistical learning based on previous
human actions to apply the appropriate strategy and autonomously solve the problem.
Our approach aims to increase the reliability of the robot system while reducing the
need for human intervention in assembly tasks.
Learning assembly operations is a time-consuming and challenging process. Therefore,
the last goal is to enable the robot to autonomously learn certain suboperations
of the assembly process. One of the obstacles that autonomous learning methods in
robotics need to overcome is the large search space that the robot has to explore before
it learns to perform the task correctly. The robot is inherently in physical contact
with the environment during the robot assembly. Although learning contact tasks is
traditionally considered more challenging because one has to consider interaction forces
between the robot system and the environment, learning physically constrained tasks
can be more efficient. The key observation is that there are fewer admissible movement
directions due to environmental constraints, reducing the number of learning
parameters and enabling faster policy learning. In that respect, we propose a novel
three-level hierarchical reinforcement learning scheme that includes a compliant controller
at the lower level, which moves safely along the constraints, the intermediate
level that systematically searches for possible states where movement is possible in
different directions, and the high-level sequence optimization.
All developed approaches have been validated on a collaborative robotic platform
and evaluated in various experiments, demonstrating the effectiveness of the proposed
methods.
|
