Using efference copy and neural control for adaptive walking on different terrains

Abstract

Event Abstract Back to Event Using efference copy and neural control for adaptive walking on different terrains Johannes Schröder-Schetelig1*, Poramate Manoonpong1 and Florentin Wörgötter1 1 BCCN Göttingen, Germany Around the mid-19th century, Holst and Mittelstaedt (1950) [1] demonstrated in animal models that motor commands are copied within the central nerve system (CNS). These copies help to distinguish 'reafference' (afference activity due to self-generated motion) from 'exafference' (afference activity due to changes in the external world). They can be also used to compare with the actual sensory feedback in order to subtract the self-generated sensation for state estimation as well as maintaining stable perception. In the early 1960s, Held (1961) [2] indicated that efference copies and the reafference cannot be directly compared due to the different dimensionality between motor commands and sensory feedback. Therefore, he proposed a neural mechanism that transforms an efference copy signal into an expected sensor input to compare to the actual incoming sensory signal. This neural transformation mechanism is known as 'internal model' [3]. Based on these biological findings, we apply such principles for adaptive walking on different terrains of our biped robot RunBot [4]. By doing so, we copy motor signals (efference copy) together with angle sensors of its leg joints. Then, they are transformed into the expected vestibular-like sensory feedback under normal walking condition (walking on a level floor) through a simple feedforward neural network (internal model). This sensory expectation is used to compare with the actual feedback. The differentiation of these two signals will control the body posture of RunBot to obtain stable walking, the result of which enables it to successfully walk on different terrains, e.g., different slopes versus a level floor.