Learning inverse dynamics models in O(n) time with LSTM networks

Inverse dynamics model learning is crucial for modern robots where analytic models cannot capture the complex dynamics of compliant actuators, elasticities, mechanical inaccuracies, frictional effects or sensor noise. However, such models are highly nonlinear and millions of samples are needed to encode a large number of motor skills. Thus, current state of the art model learning approaches like Gaussian Processes which scale exponentially with the data cannot be applied. In this work, we developed an inverse dynamics model learning approach based on a long-short-term-memory (LSTM) network with a time complexity of O(n). We evaluated the approach on a KUKA robot arm that was used in object manipulation skills with various loads. In a comparison to Gaussian Processes we show that LSTM networks achieve better prediction performances and that they can be trained on large datasets with more than 100,000 samples in a few seconds. Moreover, due to the small training batch size of for example 128 samples, the network can be continuously improved in life-long learning scenarios.