Current research in mobile robotics focuses more and more on enabling robots and humans to share a common workspace. A well known research initiative in this direction is the Factory of the Future, which has the goal to develop smart factories with networked tools, devices, and mobile manipulation platforms (e.g. the KUKA youBot). It is also known as Industry 4.0, which was coined at the Hanover Fair in 2011 (Kagermann et al. 2011).
Although robot localisation is a requirement for multi-robot collision avoidance, most approaches assume perfect sensing and positioning and avoid local methods by using global positioning via an overhead tracking camera (Alonso-Mora et al. 2015a) - or are purely simulation based (van den Berg et al. 2011). Nevertheless, to be able to correctly perform local collision avoidance in a realistic environment, a robot needs a reliable position estimation of itself and the other agents and humans without the help of external tools. Additionally, multi-robot systems in a real-world environment need methods to deal with the uncertainty in their own positions, and the positions and possible actions of the other agents.
Workspace 5 Robot Simulation Download
Lastly, we introduce the combination of the sampling based approach with the DWA method. The DWA approach is commonly used as control algorithm for local control. It is the standard method which is used on many platforms when using ROS (Quigley et al. 2009). It uses forward simulations of a set of velocity commands, known as trajectory rollouts. In our experiments, we show how the sampling based method can successfully be combined with the DWA approach to ensure good navigation within the proximity of other robots, static obstacles and humans.
This work introduces a local collision avoidance approach that deals a.o. with the problems of multiple robots sharing the same workspace with or without humans. An overview of existing (global and local) approaches for human aware navigation (Kruse et al. 2013) shows that the main focus of current research is on the comfort, naturalness and sociability of robots in human environments. This usually entails only one robot acting in a group of humans, i.e. as a personal assistant. Our approach however, is aimed at a different distribution of agents, namely many robots navigating together with many humans in the same shared workspace.
Another problem occurs when the workspace is cluttered with many robots and these robots to not move or to only move slowly. As shown Fig. 1b, the VOs are translated by the velocity of the other agents. Thus, in these cases, the apexes of the VOs are close to the origin in velocity space. Additionally, if static obstacles such as walls are included, any velocity will lead to a collision eventually, thus rendering the robots immobile. This problem can be solved using truncation.
When the velocity samples have been created, the robot uses a forward simulation to predict the effect of the given velocity in the configuration space. In other words, the robot simulates the trajectory, if the given velocity would be commanded. Afterwards, this trajectory is scored based on various cost functions. For instance, the robots footprint is imposed on each point in the simulated trajectory and if the robot is in collision at any point, the trajectory is excluded. Other cost functions are, for instance, the distance to the goal location and the distance to the given path. Figure 6 shows a graphical interpretation of the approach for a differential drive robot.
While the previous algorithms, CALU and COCALU, provide guaranteed safety and even optimality for the individual agents, there are still some limitations that remain. Specifically, the algorithms calculate the velocity that is closest to the preferred velocity and still safe. This implies that the robots always pass each other within only marginal distances. While this approach is feasible in simulation, in real world applications it is not always possible to exactly control the velocity of the robots. With only marginal distances between the robots that pass each other, there is an increased risk that the smallest error in control will lead to a collision. An additional limitation is that all agents, either human or robot, are treated in the same way, while it would be desirable to preserve more distance from humans than from other robots.
To tackle these problems, we introduce a pro-active local collision avoidance system for multi-robot systems in a shared workspace that aims to overcome the stated limitations. The robots use the velocity obstacle paradigm to choose their velocities in the input space; however, instead of choosing only the closest velocity to the preferred velocity, more cost features are introduced in order to evaluate which one is the best velocity to choose. This allows us to apply different weights or importance factors for passing humans, other robots, and static obstacles. Furthermore, we introduce a smart sampling technique that limits the need to sample throughout the whole velocity space. The resulting algorithm is decentralised with low computational complexity, such that the calculations can be performed online in real time, even on lower-end onboard computers.
In the previous section, we have shown on how to use Monte Carlo simulations to generate the samples for the velocity of the robots. As another solution, we can generate the samples according to the motion model of the robots and translate the velocities based on the motion model to an approximated holonomic speed. This idea of so-called trajectory rollouts is applied in the well know DWA-planner (Fox et al. 1997) which is commonly used in ROS. The controller generates velocities according to the dynamic motion constraints of the robots and predicts the position-based motion model of robots. Each trajectory is evaluated according to various cost functions as presented in Sect. 3.5. While these cost function are in configuration space and not in velocity space, the similarities to our COCALU with Monte Carlo sampling approach allows us to easily combine the two planners. We can use DWA to generate the velocity samples and trajectories and evaluate the configuration space based critics as the normal DWA-planner would use and then evaluate the trajectory based on our COCALU with Monte Carlo sampling cost functions. Since the trajectory is already available, the translation to velocity space is straight forward by computing the differences of the starting point and end points and dividing by the simulation time.
We have evaluated our approach in simulation using Stage (Gerkey and Mataric 2003; Vaughan 2008) and in real-world settings. Simulation allows us to investigate the system performance using many repetitions and various extreme settings, i.e. a very dense settings with a lot of robots.
CoppeliaSim is used for fast algorithm development, factory automation simulations, fast prototyping and verification, robotics related education, remote monitoring, safety double-checking, as digital twin, and much more. You can find a feature overview here.
NVIDIA RTX Desktop Manager, included with your NVIDIA RTX Enterprise/Quadro driver or downloadable as a standalone app, helps you maximize productivity by optimizing your workspace layout. It also helps you work faster through the following streamlined features and functions:
Control Buttons: The control buttons enable the control of the interface. Play button sends the code written by User to the Robot. Stop button stops the code that is currently running on the Robot. Save button saves the code on the local machine. Load button loads the code from the local machine. Reset button resets the simulation(primarily, the position of the robot).
Workspace is the space the end effector of the robot can reach. It is limited by the mechanical structure of the robot. The industrial robot providers usually provide the workspace or work range for their robot in the manual. For example, the following image is the work range of irb120 robot, the robot we use in this exercise. 2ff7e9595c
Comments