5 Conclusions

This paper illustrated an application of Locally Weighted Projection Regression to a complex robot learning task. The O(n) update complexity of LWPR, together with its statistically sound dimensionality reduction and learning rules allowed a reliable and successful real-time implementation of the algorithm on an actual anthropomorphic robot. From the viewpoint of learning control, these results demark the first time that complex internal models can be learned autonomously in real-time on sophisticated robotic devices. In an even more challenging application, we will implement LWPR for inverse dynamics learning on a 30 DOF full-body humanoid robot (Fig.1b) in the near future.