We present the control architecture of a modular robotic system designed for hospital logistics. The robotics system is an automated guided vehicle that can autonomously perform tasks like carrying and delivering objects, medicines or devices, while interacting with humans in the hospital environment. Specically, the robotic platform is designed to dock and move passive vehicles (like carts, containers, etc.), which dynamically change the robot shape and function during the task execution. We describe the overall control architecture focusing on the executive and the planning systems. We discuss the system at work in dierent scenarios considering both autonomous and interactive tasks.
Hospital logistic is an important aspect for the healthcare systems since it allows
to optimize costs and to improve the quality of service in hospitals [
In the following sections we rstly introduce the scenario in which the work is carried out, then we detail the overall control architecture focusing on its principal components and nally we describe case studies of supplies transportation in both simulated and real hospital scenarios.
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). The proposed robotic system is framed in the context of the RoMoLO - Modular robots for hospital logistics project. The project proposal aims at creating a mobile and modular robotic system capable of performing dierent logistic tasks in a hospital scenario. The proposed platform is an omnidirectional mobile robotic base able to hook passive vehicles with dierent functions: transport of sheets, drug dispensing system, telepresence vision system, etc.. The peculiarity of the system is the adoption of a single type of mobile base, which can then integrate dierent types of passive vehicles on wheels for multiple usage. Specically, an automatic coupling system can dock dierent types of passive vehicles: the mobile robot moves under the passive vehicle, docks it, and carries it through restricted environments, replacing the manual thrust. Therefore, the same mobile robot platform can be exploited for dierent logistic processes, e.g. to circulate drugs, sheets and waste in the hospital, interacting with sta and patients to receive instructions or communicate information.
These vehicles, compared to the state of the art, should allow greater exibility while containing costs since a single mobile platform may be adapted to perform several tasks. In summary, the proposed robotic system should support the following functionalities: autonomous or remote navigation, pick-carry-place dierent passive vehicles, physical human-robot interaction, possibility of manual driving by lightly pushing the vehicles and using the mobile robot as a force amplier. 3
In this section, we illustrate the overall control architecture of the RoMoLO
system (see Figure 1) describing its main components and functionalities. In
particular, we rst describe the executive system, then we focus on task and
motion planning/execution, and nally on human-robot interaction.
Executive System. The robot behavior is managed by the executive system that
interacts with the robotic system and the human operator. We deploy the
framework by [
Task and motion planning and execution. The supervisory system can also invoke
task and motion planning processes during the execution of the tasks. Dierent
task and motion planning methods can be available to the system and can be
associated to dierent contexts. The planning and execution process starts from
the executive system, where high level tasks are hierarchically decomposed into
sub-tasks at dierent levels of abstraction. The task decomposition process can
allocate path and motion planning processes. For instance, Figure 2 depicts task
and motion planning during the execution of simple cart-taking activity. Here,
the hospital map is provided, a Dijkstra-based path planning is deployed to
nd a feasible sequence of waypoints, while the navigation between waypoints
is performed using an RRT-based planner. In this case, task, path, and motion
planning are separated processes. In other cases, in the presence of more complex
activities involving multiple carts movement with dierent shapes in cluttered
environments, the system can also deploy combined task and motion planning
engines that take into account both logical and kinematic constraints (see e.g.
[
Physical human-robot interaction system. The proposed platform should also
support the human guidance (performed by hospital employees) as normal
passive carts, hence physical human-robot interaction has to be taken into account.
To this end the robotic system is endowed with seven load cells, which are
directly integrated in the lifting system as shown in Figure 3. Four of the seven
cells are positioned on the corners of the lifter allowing the robot to estimate the
weight and the center of mass of the cart (purple arrows), while the other three
are along the x and y and around the z axis of the lifter in order to estimate
the torque and the forces applied by the cart (blue, red and green arrows). This
mechanism allows the HRI module to estimate the forces due to the natural cart
shifting and to isolate only the external ones (provided by the human). When
relevant external forces are detected, a HRI behavior acquires the control of the
robot actuators in order to make the robot compliant with respect to the
human guidance. Notice that robot compliance here is needed at dierent levels of
abstraction, not only for the robot motion, but also at the path and task level.
In this respect, we are investigating how to adapt methods similar to the ones
proposed in [
In this section, we describe the RoMoLO system at work in hospital scenarios. In particular, the aim is to illustrate the following features: structured task orchestration, task and motion planning, and human-robot interaction. The testing environment is a 15x10m closed space composed by a corridor connecting 3 rooms through 3 openings of 1:2m width (Figure 2). The AGV has four independent mecanum wheels, 4 LIDARs (one for each of the robot corner) for the obstacle avoidance and a frontal camera for carts recognition using ARtags. Flexible Task Execution. The rst case study consists of an autonomous cart transportation from a room to another. The operation is performed by providing the robot with high level commands. For instance, in Figure 4 (top right) the high level task moveCart(marker4; room2; room3) is decomposed into two subtasks, takeCart and leaveCart, which in turn are decomposed by the system till the primitive actions (Figure 4 (top left)).
The initial conguration of the task is the one depicted in Figure 4 (top left). Initially, the robot is heading towards the room2 and it is reaching the Landmark corr2 rst (middle and bottom, rst left). Then a sequence of four subtasks (blue ovals in Figure 4 (top right)) is executed. The robot moves towards room2 (sutask 2). When the cart is detected by the robot (marker marker4), pickCart is enabled (subtask 3) to reach an approach point, move the robot under the cart and lift it. Once lifted the cart, takeCart is completed and leaveCart (in room3) is enabled. When the robot carries a cart, the robot’s footprint in the motionPlanner tasks is appropriately increased. Therefore, the task aects not only logic constraints, but also the kinematic ones. Figure 4 (bottom) shows how the robots size is updated when the cart is carried (third and fourth, bottom). Notice also that multiple tasks can be allocated and exibly orchestrated by the executive system. For instance, during navigation additional deliveries may be on-line requested and opportunistically executed. Notice also that the WM provides an explicit representation of the activities which are planned, executed, enabled, disabled, etc., therefore it also supports task-level explainability. The robot executive status can be inspected at dierent levels of abstraction, while a human operator can also disable or enable activities at run-time. We are currently investigation adaptive human-machine interfaces suitable for this scenario.
Manual carriage. In a dierent case study, manual carriage transport was simulated using the estimate of external forces through the load cells of the lifting system. Figure 5 shows the manual transport of the trolley from room room3 to room room2. The force impressed by the human to move the trolley is between 1 and 2 kg. During the hand guidance, the robot and the human activities are supervised and regulated by executive system, this way the robot can follow the human guidance, while maintaining a compliant behavior. A specic HRI be
Goal Cart2 placed Cart1 placed
Start
Cart1 taken
Cart2
taken
havior is activated to manage this interaction, but we are currently developing
and testing more complex interactions as in [
Task and motion planning with two carts. In this case study, we consider a simulated scenario where two carts are present. As illustrated in Figure 4 (left) the robot has to move both carts outside of the room. Notice that this task can only be performed by moving the left cart (cart1) before the right one (cart2) because the second one is blocking the room opening. In this particular case, decoupled task and motion planning can be ineective: the executive system may select the wrong cart to be moved rst and this fault will be detected during the execution, in so inducing replanning. In order to avoid these situations, we developed an integrated task and motion planner. In particular, in the proposed approach, task and motion constraints are handled by an extended RRT-based planning system, which generates not only the robot movement, but also the tasks the robot is to execute in order to reach the goal state. In the case of the two carts, the result of this combined task and motion planning process is depicted in Figure (right) where the blue rectangles represent samples of the robot plan. In the planned sequence the robot rstly takes cart1 from the room leaving it in the right-position, afterwards the cart2 can be taken and transported outside to the left-position. The integrated plan can be generated in about 3 sec (3:11 avg and 1:67 std of 30 runs on a platform Intel i5 5200U 2:20GHz, 8Gb ram with a single threaded implementation). We are currently carrying out tests in simulation to assess the performance of the integrated task and motion planner in dierent contexts. 5
We presented the control architecture of a modular robotic system designed for hospital logistics. We described the overall system that enables exible execution of multiple complex tasks. We discuss the system at work in dierent scenarios considering both autonomous and interactive tasks. In this scenario, we are currently investigating: i) human guidance in the presence of complex tasks; ii) human-machine interfaces for the logistic scenario; iii) methods for integrated task and motion planning, where both logical and kinematic constraints are considered.
Acknowledgement. The research leading to these results has been supported by the H2020-ICT-731590 REFILLs, MISE ROMOLO, and MIUR PON ICOSAF.