Figure 5 - uploaded by Sylvain Calinon
Content may be subject to copyright.
Source publication
Contexts in source publication
Context 1
... key idea is to learn a feedforward compensative ZMP Figure 15: Snapshots of short period disturbance simulation recorded from 22s and with 0.4s time interval. Upper row shows the results with learning and lower row shows the result without learning. ...
Context 2
... improved walking robustness was shown in simulations and experiments. Figure 15 depicts a simulation result showing the stability of the DLR TORO humanoid robot against external disturbances during walking with and without learning. Improved convergence speed and reduced ZMP error during transition phases between different walking motions are achieved. ...
Similar publications
Citations
... In the first test, we study the transportation from Note that all locations are described with respect to the robot base frame. In addition to the desired starting and ending locations in the transportation task, we also introduce additional position constraints which require the robot passing through two via-points (plotted by These trajectories are obtained from Calinon and Lee (2017). * * We here only consider position requirements, but velocity constraints can also be directly incorporated in desired points. ...
Imitation learning has been studied widely due to its convenient transfer of human experiences to robots. This learning approach models human demonstrations by extracting relevant motion patterns as well as adaptation to different situations. In order to address unpredicted situations such as obstacles and external perturbations, motor skills adaptation is crucial and non-trivial, particularly in dynamic or unstructured environments. In this paper, we propose to tackle this problem using a novel kernelized movement primitive (KMP) adaptation, which not only allows the robot to adapt its motor skills and meet a variety of additional task constraints arising over the course of the task, but also renders fewer open parameters unlike methods built on basis functions. Moreover, we extend our approach by introducing the concept of local frames, which represent coordinate systems of interest for tasks and could be modulated in accordance with external task parameters, endowing KMP with reliable extrapolation abilities in a broader domain. Several examples of trajectory modulations and extrapolations verify the effectiveness of our method.
... Programming a robot to use exploration motions by conventional interfaces is elaborate and requires expert knowledge. To ease this procedure, learning from demonstration [6] is employed as teaching interface to directly extract the desired behavior. Hereby, the user demonstrates the task 1 consisting of exploration and manipulation motions multiple times by kinesthetic teaching. ...
The recent generation of compliant robots enables kinesthetic teaching of novel skills by human demonstration. This enables strategies to transfer tasks to the robot in a more intuitive way than conventional programming interfaces. Programming physical interactions can be achieved by manually guiding the robot to learn the behavior from the motion and force data. To let the robot react to changes in the environment, force sensing can be used to identify constraints and act accordingly. While autonomous exploration strategies in the whole workspace are time consuming, we propose a way to learn these schemes from human demonstrations in an object targeted manner. The presented teaching strategy and the learning framework allow to generate adaptive robot behaviors relying on the robot's sense of touch in a systematically changing environment. A generated behavior consists of a hierarchical representation of skills, where haptic exploration skills are used to touch the environment with the end effector, and relative manipulation skills, which are parameterized according to previous exploration events. The effectiveness of the approach has been proven in a manipulation task, where the adaptive task structure is able to generalize to unseen object locations. The robot autonomously manipulates objects without relying on visual feedback.
... Hand programming of all these tasks is not feasible. Hence, researchers have investigated how to acquire novel tasks in an intuitive manner [1], [2]. A possible solution is to demonstrate the task to execute, for example by physically guiding the robot towards the task completion [3], [4]. ...
... This avoids jumps in the velocity but it may cause jumps 1 Institute of Robotics and Mechatronics, German Aerospace Center (DLR), Weßling, Germany matteo.saveriano@dlr.de. 2 Human-Centered Assistive Robotics, Technical University of Munich, Munich, Germany felix.franzel@tum.de, dhlee@tum.de. ...
... Given the amplitude a and the centers c i , the parameters w i are learned from demonstration using weighted least square [5]. From (2), it is clear that f p (h) vanishes for h → 0. ...
In this paper, we focus on generating complex robotic trajectories by merging sequential motion primitives. A robotic trajectory is a time series of positions and orientations ending at a desired target. Hence, we first discuss the generation of converging pose trajectories via dynamical systems, providing a rigorous stability analysis. Then, we present approaches to merge motion primitives which represent both the position and the orientation part of the motion. Developed approaches preserve the shape of each learned movement and allow for continuous transitions among succeeding motion primitives. Presented methodologies are theoretically described and experimentally evaluated, showing that it is possible to generate a smooth pose trajectory out of multiple motion primitives.
... For instance a HMM model can be incrementally refined for incorporating a physical human interaction by using a motion refinement tube [59]. The encoding motion can also be utilized for performing collaborative behaviors [20,90]. For e.g. in a collaborative object manipulation task, the robot not only learns the desired path but also the required interaction forces/torques to carry the object through the desired path. ...
Robotics researchers have been always in pursuit of developing machines that can act like humans. To achieve such an ambition, it is important to understand how humans learn. From a new born, until an adult, a human goes through several stages of developments in acquiring new skills. We do not learn every thing on our own and most of it is taught to us by our parents and teachers. For a robot to act like humans, it is also important for it to learn like humans. An important aspect of learning like humans is to be able to mimic and generalize a demonstrated skill. Programming by Demonstration (PbD) focuses on this aspect of robotics.
Multimodal density estimation is a preliminary step of the PbD approaches presented in this thesis. The main challenges in Expectation Maximization (EM) based density estimation are not knowing the original number of clusters in the data, initialization of model parameters for EM and overfitting or under fitting of the learned model. These challenges are addressed by the simulated annealing based EM approach presented in this work. The performance of the presented approach outperforms other EM based approaches in similar settings and is validated through synthetic as well as real datasets. Next I show that a learned mixture model can be utilized for PbD tasks.
With kinesthetic teaching as an input modality I present a Gaussian Mixture Model (GMM) based task parameterized skill learning approach. Existing PbD approaches require a large amount of demonstrations. Additionally they also perform poorly when generalizing beyond the demonstrated regions. The approach presented in this thesis addresses these shortcomings by presenting a new task parameterized skill learning method. As EM can utilize incomplete data, it is shown that the generalization capabilities of a learned skill can be significantly improved by introducing incomplete data spanning the task parameters i.e. the input space.
In PbD, variables of interest, which are also termed as task parameters, have to be identified and extracted during motion reproduction. To avoid extraction of task parameters, it is shown that by modelling the forcing terms of a Dynamic Movement Primitive (DMP) with a Convolution Neural Network (CNN), the high dimensional camera image can be processed directly to reproduce a learned skill. The proposed architecture possesses the desirable attributes associated with a dynamical system, as well as generalization properties associated with the deep learning.
Apart from kinesthetic teaching, this thesis also focuses on developing PbD approaches by utilizing teleoperated demonstrations. Since it is not always possible to co-locate a human teacher along with a robot, for instance in deep sea or space applications, teleoperated demonstrations have to be utilized for learning. The challenges which arise from teleoperated demonstrations are their high temporal and spatial variations, different initial and final configurations and incomplete demonstrations. This thesis addresses these challenges by proposing a simultaneous trajectories synchronization and encoding algorithm. Moreover it is shown that the learned control can also be utilized in shared teleoperation setting, by executing the learned skill in conjunction to a human operator. The PbD approaches presented in this thesis outperform the state of the art PbD
approaches devised for similar tasks in simulated as well as real robot experiments.
... The interested reader is also referred to [70], where different approaches on learning soft task priorities have been reviewed recently. ...
... The main advantage of this scheme is that smooth adaptation of task priorities is straightforward (by interpolating λ i ) at the cost of uncontrolled interference between tasks [24]- [27]. Refer to [28] for a survey of priority learning approaches. ...
Recently, several approaches have attempted to
combine motion generation and control in one loop to equip
robots with reactive behaviors, that cannot be achieved with
traditional time-indexed tracking controllers. These approaches
however mainly focused on positions, neglecting the orientation
part which can be crucial to many tasks e.g. screwing. In
this work, we propose a control algorithm that adapts the
robot’s rotational motion and impedance in a closed-loop
manner. Given a first-order Dynamical System representing
an orientation motion plan and a desired rotational stiffness
profile, our approach enables the robot to follow the reference
motion with an interactive behavior specified by the desired
stiffness, while always being aware of the current orientation,
represented as a Unit Quaternion (UQ). We rely on the Lie
algebra to formulate our algorithm, since unlike positions, UQ
feature constraints that should be respected in the devised
controller. We validate our proposed approach in multiple robot
experiments, showcasing the ability of our controller to follow
complex orientation profiles, react safely to perturbations, and
fulfill physical interaction tasks.
Learning and reproducing the tracked robot’s demonstration trajectory is a promising intelligent path planning solution that can summarize and extract the relevant characteristics of the tracked robot’s desired trajectory. However, the existing methods are difficult to ensure the robustness when there are deviations in the demonstration trajectory data or when the task constraints are added. In order to address the problem, this paper proposes a novel trajectory learning and reproduction method for tracked robots that is based on Bagging algorithm and GMM/HSMM. The Bagging algorithm is used to randomly resample the demonstration trajectory data and construct the sub trajectory data sets. Then, GMM/HSMM is used to train and learn these sub trajectory data sets, and the output probability density function of the hidden state and the Gaussian component with the largest average position probability are selected as the motion elements. According to the mean and variance of these motion elements, the least square method is used to reconstruct the tracked robot’s trajectory. On the basis of using the IPSO algorithm to optimize the position of constraint points in each sub trajectory data sets, combined with the weight method, the learning results are integrated to realize the trajectory reproduction of a tracked robot under task constraints. The final results and analysis show that the proposed method can successfully realize the trajectory learning and reproduction of tracked robot, as well as ensure that the reproduced trajectory can pass through the required task constraint points without increasing the algorithm’s complexity.
