How to Get from Interpolated Keyframes to Neural Attractor Landscapes and Why.

摘要

Abstract — Storing poses of a humanoid robot as keyframes and interpolating between them is a common technique used to produce robot motion. With this technique complex motion sequences can be recorded and modified easily by hand, but it is difficult to incorporate sensory feedback to stabilize the robot’s trajectories. Neural Networks, in contrast, are suitable for sensorimotor loops, but it is hard to design predefined attractor shapes with explicit timing constraints. We introduce basic neural building blocks and show how to interconnect and parameterize them in order to achieve desired motion sequences. We explain why a purely neural approach may be superior to hybrid control architectures when using sensory feedback, especially if one wishes to make the robot’s motions more robust by artificial evolution. Index Terms — Robot Motion, Keyframes, Neural One-Shot I.