Estimating Body Pitch from Distributed Proprioception in a Hexapod

Abstract

Adaptability of legged locomotion relies on distributed proprioceptive feedback from the legs. Apart from low-level control of leg movements, proprioceptive cues may also be integrated to estimate overall locomotion parameters relevant to high-level control of behavior. For example, this could be relevant for reliable estimates of body inclination relative to the substrate, particularly in animals that lack dedicated graviceptors such as statocycsts. With regard to robotic systems, distributed proprioception could exploit physical interaction with the substrate to improve the robustness of inclination estimates. In insect locomotion, it is unknown how overall parameters such as body inclination or forward velocity may be represented in the nervous system. If proprioceptive encoding was optimal, the afferent activity pattern of distributed proprioceptive cues from across the body should be a suitable representation in itself. However, given noisy encoding in multiple afferent spike trains, it is unknown (i) how reliable the parameter estimates can be, and (ii) which parts of a distributed proprioceptive code are most relevant.

Here we use a database on unrestrained whole-body kinematics of walking and climbing stick insects in conjunction with simple spiking proprioceptor models to transform sets of joint angle time courses into corresponding sets of spike trains. In total, we tested four different types of models: a reference model without proprioceptive encoding and three proprioception models with different filter properties and spike generators. Within each model, we compared 4 × 4 conditions that differed in number and combination of joints and legs. Our results show that the contribution of middle and hind legs is of similar relevance for the estimation of body pitch, whereas front legs contribute only very little. Furthermore, femoral levation proved to be the most relevant degree of freedom, whereas estimates based on protraction and extension angles were less accurate.