Data-Efficiency

I adopt the model-based RL approach as the primary means for achieving data-efficiency. My work so far is limited to only model learning. Policy search with a learned model is planned for the future.

 

Learning dynamics models as hybrid automata

In this work, the goal is to investigate whether learning forward dynamics models with the structure of hybrid automata can lead to better data-efficiency. Hybrid automata is ideal for representing the peculiarities of contact dynamics. 

Khader, S. A., Yin, H., Falco, P., & Kragic, D. (2020). Data-efficient model learning and prediction for contact-rich manipulation tasks. IEEE Robotics and Automation Letters. [IEEE] [arXiv]

 

Evaluation of iLQR with learned dynamics model

In this work, the popular method, iLQR with learned dynamics model, is evaluated for an assembly task. The convergence of the method, to a successful policy, in only 160 trials validates the data-efficiency of the model-based RL approach.