Filters








12 Hits in 7.0 sec

Bayesian Optimization Using Domain Knowledge on the ATRIAS Biped [article]

Akshara Rai, Rika Antonova, Seungmoon Song, William Martin, Hartmut Geyer, Christopher G. Atkeson
2017 arXiv   pre-print
We aim to overcome this problem by incorporating domain knowledge to reduce dimensionality in a meaningful way, with a focus on bipedal locomotion.  ...  In this paper, we present a generalized feature transform applicable to non-humanoid robot morphologies and evaluate it on the ATRIAS bipedal robot -- in simulation and on hardware.  ...  Utilizing trajectory data with Behaviour-based Kernels Bayesian Optimization has been used to optimize controllers in various robotics domains.  ... 
arXiv:1709.06047v1 fatcat:kvywd4e5avaz5dxxryltc4ifyq

Using Simulation to Improve Sample-Efficiency of Bayesian Optimization for Bipedal Robots [article]

Akshara Rai, Rika Antonova, Franziska Meier, Christopher G. Atkeson
2018 arXiv   pre-print
To answer this, we create increasingly approximate simulators and study the effect of increasing simulation-hardware mismatch on the performance of Bayesian optimization.  ...  Experiments on the ATRIAS robot hardware and another bipedal robot simulation show that our approach succeeds at sample-efficiently learning controllers for multiple robots.  ...  Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding organizations.  ... 
arXiv:1805.02732v1 fatcat:odaqdykfpbhfvnosgl4cspz6hq

Deep Kernels for Optimizing Locomotion Controllers [article]

Rika Antonova, Akshara Rai, Christopher G. Atkeson
2017 arXiv   pre-print
First, we demonstrate improvement in sample efficiency when optimizing a 5-dimensional controller on the ATRIAS robot hardware.  ...  To address this, prior work has proposed using domain expertise for constructing custom distance metrics for locomotion. In this work we show how to learn such a distance metric automatically.  ...  Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding organizations.  ... 
arXiv:1707.09062v2 fatcat:nbmemaqzszclpmejcqyzkvlwxi

Using Deep Reinforcement Learning to Learn High-Level Policies on the ATRIAS Biped [article]

Tianyu Li, Akshara Rai, Hartmut Geyer, Christopher G. Atkeson
2018 arXiv   pre-print
We present our results on an ATRIAS robot and explore the effect of action spaces and cost functions on the rate of transfer between simulation and hardware.  ...  However, domain randomization can make the problem of finding stable controllers even more challenging, especially for underactuated bipedal robots.  ...  Recent work learns parameters of expert controllers on hardware sample-efficiently, for example [9] and [10] use Bayesian Optimization.  ... 
arXiv:1809.10811v1 fatcat:yjgzk6qfvfaz5mzmv4xogl7q6e

Learning Fast Adaptation with Meta Strategy Optimization [article]

Wenhao Yu, Jie Tan, Yunfei Bai, Erwin Coumans, Sehoon Ha
2020 arXiv   pre-print
The ability to walk in new scenarios is a key milestone on the path toward real-world applications of legged robots.  ...  The key idea behind MSO is to expose the same adaptation process, Strategy Optimization (SO), to both the training and testing phases.  ...  ACKNOWLEDGMENT The authors gratefully thank Tingnan Zhang, Karol Hausman, Benjamin Eysenbach, the locomotion team at Google Robotics, and the anonymous reviewers for valuable discussion and suggestions  ... 
arXiv:1909.12995v2 fatcat:3kndegpgjfe6nbqq4azmmgbr7u

A survey on policy search algorithms for learning robot controllers in a handful of trials [article]

Konstantinos Chatzilygeroudis, Vassilis Vassiliades, Freek Stulp, Sylvain Calinon, Jean-Baptiste Mouret
2019 arXiv   pre-print
A second strategy is to create data-driven surrogate models of the expected reward (e.g., Bayesian optimization) or the dynamical model (e.g., model-based policy search), so that the policy optimizer queries  ...  We show that a first strategy is to leverage prior knowledge on the policy structure (e.g., dynamic movement primitives), on the policy parameters (e.g., demonstrations), or on the dynamics (e.g., simulators  ...  Helmholtz Association through the project "Reduced Complexity Models"; the European Commission through the projects H2020 AnDy (GA no. 731540) and MEMMO (GA no. 780684); the CHIST-ERA project "HEAP";  ... 
arXiv:1807.02303v5 fatcat:df5wuzgfp5aorisx757hr655bi

Learning to Learn from Simulation: Using Simulations to Learn Faster on Robots

Akshara Rai
2019
Our hardware experiments on the ATRIAS robot, and simulation experiments on a 7-link biped model, show that these feature transforms capture important aspects of walking and accelerate learning on hardwareand  ...  this transformation, hand-designed features based on knowledge of human walking and using neural networks to extract this information automatically.  ...  ATRIAS biped and 7-link bipedal robot.  ... 
doi:10.1184/r1/7707965 fatcat:4xvw3ntwnzhg5f72nd3gjcqy3i

A Survey on Policy Search Algorithms for Learning Robot Controllers in a Handful of Trials

Konstantinos Chatzilygeroudis, Vassilis Vassiliades, Freek Stulp, Sylvain Calinon, Jean-Baptiste Mouret
2019 IEEE Transactions on robotics  
A second strategy is to create data-driven surrogate models of the expected reward (e.g., Bayesian optimization) or the dynamical model (e.g., model-based PS), so that the policy optimizer queries the  ...  In this article, we show that a first strategy is to leverage prior knowledge on the policy structure (e.g., dynamic movement primitives), on the policy parameters (e.g., demonstrations), or on the dynamics  ...  [81] used domain knowledge for bipedal robots (i.e., determinants of gait (DoG) [83] ) to produce a kernel that encodes the differences in walking gaits rather than the Euclidean distance of the policy  ... 
doi:10.1109/tro.2019.2958211 fatcat:kekbwy2v4fbllfd6xytm5ry3ue

Optimal Control of Compliant Bipedal Gaits and Their Implementation on Robot Hardware

William Martin
2019
We show that this control methodology leads to stable locomotion across several different gaits on the ATRIAS biped robot.  ...  Recently, researchers have proposed using the spring mass model as a compliant locomotion paradigm to create unified controllers for walking and running on bipedal systems.  ...  optimal gait plans on the ATRIAS biped.  ... 
doi:10.1184/r1/8397743 fatcat:7d27xw6j4nguncoa5gxq7po44q

Design and Evaluation of Robust Control Methods for Robotic Transfemoral Prostheses

Nitish Thatte
2019
We also propose a pair of optimization methods that allow us to select prosthesis control parameters using qualitative preference feedback from the user.  ...  In the first, we use information from an inertial measurement unit and a LIDAR distance sensor to estimate the position, orientation and fu [...]  ...  The method deals with high dimensional optimization problems by incorporating domain knowledge in the form of an offline optimization step.  ... 
doi:10.1184/r1/8397551 fatcat:ouzitvlnqfa2zgudsuwjpz26ha

Viability in State-Action Space: Connecting Morphology, Control, and Learning [article]

Steve Heim, Universitaet Tuebingen, Universitaet Tuebingen, Badri-Spröwitz, Alexander (Dr.)
2020
The main contributions of this dissertation are based on viability theory.  ...  This demonstration highlights the importance of robustness to failures for learning control: not only can failures cause damage, but they typically do not provide useful gradient information for the learning  ...  In particular, we appreciate the frequent and insightful discussions with Matthew Millard, Brent Gillespie and Andrea del Prete, as well as Friedrich Solowjow's advice on mathematical notation.  ... 
doi:10.15496/publikation-42525 fatcat:wsxpclt2mfa5tnnrximo3kezum

On the discretisation of actuation in locomotion: Impulse- and shape-based modelling for hopping robots

Fabio Felice Giardina, Apollo-University Of Cambridge Repository, Apollo-University Of Cambridge Repository, Fumiya Iida
2018
There are many reasons for robots' inferior performance, but arguably the most important one is our missing understanding of complexity.  ...  The reduction of complexity of the model equations reveals the underlying physics of the locomotion process, and we identify the importance of robot shape and mass distribution for the locomotion performance  ...  He uses a policy gradient reinforcement learning algorithm on a mechanical biped that is based on a passive dynamic walker.  ... 
doi:10.17863/cam.22672 fatcat:7s5ahmeawfcmha5i4kaninvyje