TuneNet: One-Shot Residual Tuning for System Identification and\n Sim-to-Real Robot Task Transfer

Abstract

As researchers teach robots to perform more and more complex tasks, the need\nfor realistic simulation environments is growing. Existing techniques for\nclosing the reality gap by approximating real-world physics often require\nextensive real world data and/or thousands of simulation samples. This paper\npresents TuneNet, a new machine learning-based method to directly tune the\nparameters of one model to match another using an *iterative residual tuning*\ntechnique. TuneNet estimates the parameter difference between two models using\na single observation from the target and minimal simulation, allowing rapid,\naccurate and sample-efficient parameter estimation. The system can be trained\nvia supervised learning over an auto-generated simulated dataset. We show that\nTuneNet can perform system identification, even when the true parameter values\nlie well outside the distribution seen during training, and demonstrate that\nsimulators tuned with TuneNet outperform existing techniques for predicting\nrigid body motion. Finally, we show that our method can estimate real-world\nparameter values, allowing a robot to perform sim-to-real task transfer on a\ndynamic manipulation task unseen during training. Code and videos are available\nonline at http://bit.ly/2lf1bAw.\n