Learning State-Dependent Losses for Inverse Dynamics Learning
This work addresses data efficiency in model-based control for object manipulation, offering incremental improvements in adaptation speed.
The paper tackled the problem of fast adaptation in inverse dynamics learning for model-based control by proposing meta-learned, state-dependent loss functions, which improved online adaptation speed in both simulation and real hardware tasks compared to standard loss functions.
Being able to quickly adapt to changes in dynamics is paramount in model-based control for object manipulation tasks. In order to influence fast adaptation of the inverse dynamics model's parameters, data efficiency is crucial. Given observed data, a key element to how an optimizer updates model parameters is the loss function. In this work, we propose to apply meta-learning to learn structured, state-dependent loss functions during a meta-training phase. We then replace standard losses with our learned losses during online adaptation tasks. We evaluate our proposed approach on inverse dynamics learning tasks, both in simulation and on real hardware data. In both settings, the structured and state-dependent learned losses improve online adaptation speed, when compared to standard, state-independent loss functions.