|Ph.D. Thesis Juan A. Corrales|
Title: Safe Human-Robot Interaction Based on Multi-sensor Fusion and Dexterous Manipulation Planning
Author: Juan Antonio Corrales Ramón
Fernando Torres Medina
Francisco Andrés Candelas Herías
Presentation Date: 21-07-2011
Prof. Miguel Angel Salichs Sánchez-Caballero, University Carlos III, Madrid, Spain.
Dr. Jorge Pomares Baeza, University of Alicante, Alicante, Spain.
Prof. Alfonso García Cerezo, University of Málaga, Málaga, Spain.
Prof. Véronique Perdereau, University Pierre et Marie Curie, Paris, France.
Dr. Óscar Reinoso García, University Miguel Hernández, Elche, Spain.
The main goal of this thesis is to develop human-robot interaction tasks where human operators and robotic manipulators can cooperate in the same workspace. In order to do so, this thesis proposes several techniques to fulfill two requirements which are essential for this human-robot interaction: the guarantee of the human safety during human-robot interaction and the development of dexterous manipulation which improves the skills of robotic manipulators with robotic hands.
For the first requirement, this thesis develops new safety strategies which stop the normal behavior of the robot and activate a special safety behavior when the human-robot distance is below a safety threshold. A precise full-body human tracking system is required in order to compute on real-time this human-robot distance. This thesis makes a comparison of the current technologies of human motion capture systems and proposes to use an inertial motion capture system to track the human operator who collaborates with the robot. Although the relative rotational measurements of this system are very accurate and are applied over a skeleton; the global position of this skeleton accumulates errors. Thus, this thesis uses an additional localization system based on UWB (Ultra-wideband) pulses in order to correct this error. This thesis presents three novel fusion algorithms based on Bayesian filtering in order to combine the global position measurements of both systems. The first algorithm is based on a Kalman filter, the second algorithm is based on a particle filter and the third algorithm is based on a combination of a Kalman filter and a particle filter. All these algorithms have a new structure based on recalculating the transformation matrix between the coordinate frames of both systems each time a new UWB measurement is received. This structure improves the computational cost of the proposed algorithms in comparison with previous similar approaches because it considers the complementariness between the features of both systems: the rotational accuracy and the high sampling rate of the inertial system with the positional accuracy and the low sampling rate of the UWB system.
This thesis also proposes to cover the bones which compose the skeleton registered by the previous human tracking system by a hierarchy of bounding volumes in order to obtain a precise and efficient approximation of the human-robot distance. This hierarchy is composed by three levels which contain different bounding volumes depending on the required level of detail. The first level is composed by general AABBs (Axis-Aligned Bounding Boxes) which cover with one bounding volume all the body of the human and all the body of the robot. The second level is composed by local AABBs which cover the main limbs/links of the human and the robot. The third level is composed by SSLs (Swept-Sphere Lines) which cover each bone of their skeletons. Thereby, when the human and the robot are far away from each other, only the first level is needed to compute a good approximation of the human-robot distance. Nevertheless, when they are close to each other, the second or even the third levels are required. This hierarchical representation not only improves the computational cost of the human-robot distance calculation by reducing the number of pairwise distance tests between bounding volumes, but it also improves the approximation of the human and robot surfaces by using the SSLs of the third level of the hierarchy in comparison with previous human-robot interaction systems based on spherical bounding volumes. In addition, the radii of these SSLs are modified dynamically according to the linear velocity of the bone which is covered by each SSL. This variation supposes an increase of the human safety since the bounding volumes of the third level of the hierarchy not only represent an approximation of the surface of the human and the robot bodies but also an estimation of their displacement between two execution steps of the safety strategy. This hierarchy of dynamic bounding volumes is used in safety strategies which have been applied successfully in several real human-robot interaction tasks.
For the second requirement of flexible human-robot interaction tasks, this thesis proposes a novel dexterous manipulation planner based on in-hand movements generated by multi-fingered robotic hands installed at the end-effector of robotic manipulators which cooperate with human operators. This planner receives as input an initial grasp of an object and a desired final configuration (i.e. position and orientation) of the object. This planner computes the movements of the fingers which are required to drive the object to the final configuration. It is built with a modular structure based on two levels: global and local planner. The global planner generates a trajectory of the object composed by intermediate configurations between the initial and final configurations. The local planner applies a novel contact evolution model based on a triangle mesh representation of the object and fingers surfaces in order to determine the possible evolution of the contacting primitives (i.e. vertices, edges and faces of the triangle meshes). This model provides a more general representation of the object and fingers surfaces than previous manipulation planners based on planar or parametric models. Afterwards, the local planner generates movements of the fingers which cause this contact evolution by applying the pseudo-inverse of the Jacobian matrix of each finger. Furthermore, this thesis adds an additional finger readjustment algorithm after each execution of the local planner in order to guarantee that the contacting fingers apply enough pressure to the object so that unstable grasps are avoided. This planner has been successfully applied in several real experiments of manipulation tasks with a three-fingered robotic hand equipped with tactile sensors.
The complete documentation of this thesis can be downloaded in the following links:
This section contains all the videos which are referenced in the text of this thesis. They are organized according to the chapters of this thesis:
This video shows a human-robot interaction task where a human operator takes an object from a storage box which is outside the robot’s workspace and gives this object to a robotic manipulator. This video not only contains the real sequence of the experiment but also a 3D representation of the skeleton of the human. More information can be read at Section 220.127.116.11 of this thesis.
This video shows a human-robot interaction task where a human operator cooperates with a robotic manipulator in the disassembly of a connector from a metallic structure. This video shows the real sequence of the experiment. More information about this experiment and the data obtained from the human tracking system can be read at Section 18.104.22.168 of this thesis.
This video shows an execution of a task which involves changing a blown light bulb of a streetlamp with a robotic manipulator. This video shows the original trajectory which is executed by the robot in order to leave the blown light bulb inside a storage box when there is no human intervention. More information can be read at Section 4.6.1 of this thesis.
This video shows an execution of a task which involves changing a blown light bulb of a streetlamp with a robotic manipulator. This video shows how the original trajectory of the robot is changed because of the safety behavior activation when the human is near the robot. More information can be read at Section 4.6.1 of this thesis.
This video shows an execution of an assembly task of a metallic structure where a human operator collaborates with two robotic manipulators. This video not only shows the sequence of the real experiment but also a 3D representation of the dynamic bounding volumes which are generated in the third level of the hierarchy of bounding volumes. More information can be read at Section 4.6.3 of this thesis.
This video shows a simulation of a dexterous manipulation task of a cube by the Shadow hand. This task involves a displacement of the cube by 20mm in the X axis. More information can be read at Section 5.7.1 of this thesis.
This video shows a simulation of a dexterous manipulation task of a hexagonal prism by the Shadow hand. This task involves a rotation of the prism by 5° around the Y axis. More information can be read at Section 5.7.1 of this thesis.
This video shows a simulation of a dexterous manipulation task of an icosphere by the Shadow hand. This task involves a rotation of the sphere by 5° around the Z axis. More information can be read at Section 5.7.1 of this thesis.
This video shows a dexterous manipulation task of a cylinder by a Barrett hand. This task ends when the maximum pressure of finger 3 is below the safety threshold. This video shows two different views of the real task and three plots. More information can be read at Section 5.7.2 of this thesis.
This video shows a dexterous manipulation task of a sphere by a Barrett hand. This task ends when the maximum pressure of finger 2 is below the safety threshold. This video shows two different views of the real task and three plots. More information can be read at Section 5.7.2 of this thesis.
This video shows a dexterous manipulation task of a sphere by a Barrett hand. This tasks ends when the object reaches the desired final configuration. This video shows two different views of the real task and three plots. More information can be read at Section 5.7.2 of this thesis.