Few-Shot Model-Based Adaptation in Noisy Conditions

Few-shot adaptation is a challenging problem in the context of simulation-to-real transfer robotics, requiring safe and informative data collection. In physical systems, additional challenge may be posed by domain noise, which present virtually all real-world applications. this letter, we propose to perform few-shot dynamics models noisy conditions using an uncertainty-aware Kalman filter-based neural network architecture. We show that proposed method, explicitly addresses improves error over blackbox LSTM baseline, model-free on-policy reinforcement learning approach, tries learn adaptable policy at same time. The method also allows for system analysis analyzing hidden states model during after adaptation.