- author = "Seungsu Kim",
- title = "Rapid and reactive robot control framework for catching objects in flight",
- school = "Ecole Polytechnique Federale de Lausanne",
- year = "2014",
- address = "Switzerland",
- month = "13 " # feb,
- keywords = "genetic algorithms, genetic programming, robot catching, imitation learning, reachable-space, graspable space",
-
video_url = "
https://www.youtube.com/watch?v=24TCUZISIdU",
-
URL = "https://infoscience.epfl.ch/record/196452",
-
URL = "
http://lasa.epfl.ch/publications/uploadedFiles/EPFL_TH6094.pdf",
-
DOI = "
doi:10.5075/epfl-thesis-6094",
- size = "134 pages",
-
abstract = "Humans can react extremely rapidly in the face of
unexpected changes in the environment. This is best
illustrated in sports, when tennis players run and
return a fast ball flying at a speed of around 73
meters per second. Robots, on the other hand, remain
slow and clumsy to adapt to perturbations, despite the
fact that computers process information orders of
magnitude faster than the human brain. This thesis
targets the design of controllers to endow robot with
extremely fast and appropriate reactivity in the face
of unforeseen changes in the environment. As a
benchmark, we chose the challenging task of catching
objects in flight. To react appropriately to
perturbations requires the ability to detect and
predict the effect of the observed changes in the
environment and to adapt its control plan
adequately.
This thesis, hence, addresses first the problem of predicting accurately the flying trajectory of the object. We propose a model-free method to estimate the dynamics of free-flying objects. We take a realistic perspective to the problem and investigate tracking accurately and very rapidly the trajectory and orientation of an object so as to catch it in flight. We consider the dynamics of complex objects where the grasping point is not located at the centre of mass, without having any prior information on the physical properties of the object. We also consider the dynamics of non-rigid object (such as a half-filled bottle). It is challenging as inertial properties of the object are not even constant and may change during flight. To achieve this, a density estimate of the translational and rotational acceleration is built based on the trajectories of various examples by using a machine learning approach. The estimated model of the object's dynamics is a closed form solution, and it is used in conjunction with an Extended Kalman Filter for robust tracking in the face of noisy sensing. We validate the approach for real-time motion tracking of 5 daily life objects with complex dynamics (a ball, a fully-filled bottle, a half-filled bottle, a hammer and a ping pong racket).",
-
abstract = "In the second part of the thesis, we propose a novel
methodology for learning when and where to grasp the
flying object. We develop a data-driven probabilistic
approach to estimate a distribution of admissible
grasping postures on the object and to compute the
robot's reachable space. The robot is thus able to
determine the feasible catching arm configuration
(mid-flight intercept point) along the predicted object
trajectory in real-time. The method is extended first
to compute feasible bi-manual grasping posture, in a
second stage, to determine whole-body catching posture
of a 29 degrees of freedom (DOF) humanoid robot
(excluding the fingers and neck). Finally, we show that
the models of reachable space which we developed for
our robotic framework can explain observed preference
in posture in human catching motions.
In the last part of the thesis, we address the issue of adapting on the fly the robot's arm motion so as to catch the flying object on time. We adopt a dynamical system (DS) approach to control simultaneously and in coordination the motion of the arm and fingers, so that the fingers close on time on the object. Additionally, we propose a system to synchronise the robot's motion with that of the fast moving objects, while benefiting from all the robustness properties deriving from the DS. Furthermore, we propose a generalised human-like inverse kinematics solution, by modelling human-like characteristics (the degree of torso orientation and elbow elevation) from human demonstrations and by applying the model to a generalised inverse kinematic problem. The humanoid robot is thus able to increase the human-likeness while it executes the trained task-space motion. We have validated the methods developed in the thesis, in simulation and real-world experiment with different robot platforms, iCub (53 DOF) and COMAN (29 DOF) humanoid robots and KUKA LWR robot arm (7 DOF). In particular, we demonstrated the extremely fast speed of our method in an impressive demonstration, whereby the KUKA LWR robot arm catches in-flight different objects with uneven mass distribution, such as a tennis racket and a bottle partly filled with water. We believe that our methods significantly advances the field, in offering an example of ultra-fast control in the face of uncertainty.",
-
notes = "In English. Supervisor Prof. Aude Billard, directrice
de these.
Suisse Thesis Number 6094 (2014) EPFL Docteur es sciences. GPTIPS",
