Safety in Data-Driven Controllers
1. Safe Data-Driven Contact-Rich Manipulation
In this project we address the safety of data-driven control for contact-rich manipulation. We propose to restrict the controller's action space to keep the system in a set of safe states. In the absence of an analytical model, we show how Gaussian Processes (GP) can be used to approximate safe sets. We disable inputs for which the predicted states are likely to be unsafe using the GP. Furthermore, we show how locally designed feedback controllers can be used to improve the execution precision in the presence of modelling errors. We demonstrate the benefits of our method on a pushing task with a variety of dynamics, by using known and unknown surfaces and different object loads.
Our results illustrate that the proposed approach significantly improves the performance and safety of the baseline controller.
More information can be found in:
I. Mitsioni*, P. Tajvar*, D. Kragic, J. Tumova and C. Pek, "Safe Data-Driven Contact-Rich Manipulation", 2020 IEEE-RAS 20th International Conference on Humanoid Robots, Munich, Germany, 2021
* the authors contributed equally
2. Safe Data-Driven Model Predictive Control of Systems with Complex Dynamics
In this work, we take inspiration from unsupervised outlier detection to detect unsafe behaviors. To this end, we exploit that collected datasets to train the dynamics model of a Data-Driven Model Predictive Controller (DD-MPC), are chosen to display good enough performance (e.g., as evaluated by a human worker) and encode information about safe motion examples in their dynamics. Given a clustering of the collected trajectories, we can detect the regions of high sample density that correspond to safe states, as well as the outliers which will correspond to unsafe/undesired states. This information can then be used within the DD-MPC to avoid unsafe states and drive the system towards the safe ones. We increase the informativeness of the data collection process by incorporating short intervals of (bounded magnitude) random actions in predefined controllers. Finally, we address myopic predictions by incorporating information about the long-term dynamics in the controller.
More information to come.
