Abstract
Producing feasible motions for highly redundant robots, such as humanoids, is a complicated and high-dimensional problem. Model-based whole-body control of such robots can generate complex dynamic behaviors through the simultaneous execution of multiple tasks. Unfortunately, tasks are generally planned without close consideration for the underlying controller being used, or the other tasks being executed, and are often infeasible when executed on the robot. Consequently, there is no guarantee that the motion will be accomplished. In this work, we develop a proof-of-concept optimization loop which automatically improves task feasibility using model-free policy search in conjunction with model-based whole-body control. This combination allows problems to be solved, which would be otherwise intractable using simply one or the other. Through experiments on both the simulated and real iCub humanoid robot, we show that by optimizing task feasibility, initially infeasible complex dynamic motions can be realized—specifically, a sit-to-stand transition. These experiments can be viewed in the accompanying Video S1.
Highlights
Redundant robots, such as humanoids, have enormous potential industrial and commercial utility
We define a task feasibility cost function which, when optimized over a set of concurrent tasks, results in improved trajectory from the robot. We show that this task feasibility cost function can be optimized in very few iterations using a Bayesian Optimization technique where the minimization of the acquisition function is performed using the CMA-ES algorithm. We evaluate this method in simulation and on the real iCub humanoid robot and show through a simple proof-of-concept experiment that it can result in significant improvement of the generated trajectories in a way that is practical for real world learning
After 24 rollouts in simulation, the task feasibility optimization converges to a policy which produces a successful sit-to-stand transition in both simulation and on the real robot
Summary
Redundant robots, such as humanoids, have enormous potential industrial and commercial utility. Producing feasible and useful behaviors on real robots is a challenging undertaking, when the robot must interact with the environment. This is caused, in large part, by the fact that there are always errors between what is planned, or simulated, and what is executed on a real robot due to modeling errors and perturbations. Even the most sample efficient end-to-end learning methods (e.g., Gu et al, 2016) would fail because training a model on a real robot would require an unrealistic number of evaluations, or rollouts. We show that by combining control and learning techniques, we can create low-dimensional highlevel abstractions of whole-body behaviors and efficiently correct initially infeasible motions on real robots
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
