Certified safe, fast, and real-time robot control in workspaces shared by humans and robots

Autonomous robotic manipulators that work alongside humans can bring the
production line to a new level of flexibility and efficiency and, as such, have
the potential to fundamentally transform our industry and economy. However,
despite significant progress in robotics over the last few decades, the safety
issue of robots working nearby and without barriers with human operators
has become dominant. Existing solutions and safety standards make these
collaborative robots safe by making them slower, dramatically reducing their
productivity and ultimately jeopardizing the economic attractiveness of a
collaborative workstation. This is especially true since a significant part of
the tasks in today’s industry consists of handling and moving objects. These
operations do not add value to a product and should therefore be performed
as quickly as possible.

The main objective of this PhD dissertation is to investigate whether the
performance of a robotic manipulator can be improved with real-time control
methodologies while certifying safety in cluttered and dynamic environments
and, as such, boost the productivity of collaborative workstations.

Collaborative robotic manipulators require computationally efficient and
reactive control schemes that guarantee safety, i.e., enforce a variety of state
and input constraints of nonlinear systems, are asymptotically stable, ensure
global navigation in cluttered environments, and entail a mapping between
joint and task space.

The Explicit Reference Governor (ERG) constrained control algorithm meets
many of these requirements. Therefore, I have extended and specialized the
trajectory-based ERG theory and its two main components, the Navigation
Field (NF) and the Dynamic Safety Margin (DSM), to deal in real-time with a generic robotic manipulator: a complex dynamic system subject to
actuator constraints, joint angle and velocity limitations, Cartesian end-effector
velocity constraints, and possibly time-varying primitive shaped obstacles in
the environment. I have validated the developed control law on the Franka
Emika Panda robot, showing stability, convergence, and constraint enforcement
in a collaborative work environment where the human was detected via a
motion capture system. In static environments, it is possible to guarantee
the absence of collisions. In dynamic Human-Robot Collaboration (HRC)
scenarios, it is impossible to guarantee the absence of collisions with a human
since a collision can happen due to the human hitting the robot. Instead, I
proved that, whenever a collision is imminent, the robot will become compliant
and will satisfy the Power and Force Limiting (PFL) criterion. Although the
experiments showed a very reactive and safe robot behavior, it could get stuck
in local minima since global convergence was not guaranteed.

To integrate the ERG, which guarantees constraint satisfaction in highly
dynamic environments but often gets trapped in local minima, with the Rapidly-exploring Random Tree (RRT) sampling-based motion planner, which can
solve complex path planning problems in cluttered environments but has issues
handling dynamic constraints, I designed a reference selector that sequentially
assigns one of the way-points generated by the RRT as a target reference for
the ERG. Thereby, global convergence can be guaranteed if a feasible path
can be found towards a steady-state admissible reference. I have validated this
planning and control framework in different cluttered HRC environments and
employed an off the shelf depth camera to generate an occupancy map of the
static obstacles and track human skeletons.

To make the proposed planning and control framework that enables fast
and safe HRC reproducible for the research community, I have developed a
modular ROS-based system architecture that exploits the multicore processor
capabilities and builds further on existing well-maintained ROS packages and
libraries. The framework will be open sourced on GitHub and is accompanied
by documentation provided in an online tutorial, which will enable researchers
to test the proposed framework with exemplary numerical and experimental
validation cases on the Franka Emika Panda robot, and in realistic HRC
scenarios.