From robots to brain circuits: Learning to harness elasticity for optimal motions

Abstract

Whenever a runner strikes the ground, the muscles and tendons deflect like springs and reuse the stored energy to push the athlete off the ground, thereby saving up to 84 % of the muscular energy consumption. In bionic research, roboticists increasingly replicate the elastic properties underlying the human motor performance. But the control of efficient elastic movements in changing environments is largely unknown in both robotics and neuroscience. Bionic approaches to this open question suffer from the difficult identification of functionally distinct circuits within the network of our 86 billion entangled neurons. The present dissertation fully reverses the bionic approach to explain how the human brain optimizes muscular forces under fast-changing conditions. For this endeavor, the highly interdisciplinary research first answers how the elastic dynamics of robots can be optimally harnessed and then discovers an analogous brain circuit in humans. Robotic simulations demonstrated that a fast, model-free controller coordinates multiple elastic joints as energy-efficiently as a slow, model-based optimal controller. In robotic experiments, the controller increased the amplitude of jumping by up to 67 %. While an analogous functionality would give the brain a substantial evolutionary benefit, it requires a neuronal mechanism that violates a fundamental neuroscientific consensus: that synapses can only adapt to information that is locally encoded by the pre- or postsynaptic neuron. Here, the consensus is refuted by a novel, experimentally verified model of a non-local adaptation mechanism. For its development, the robotic controller was used as a blueprint to rigorously unite scattered findings from experimental neuroscience and machine learning. In the resulting model, sensory input triggers the release of serotonin onto motor neurons. The serotonin modulates the output forces according to the same algorithm as the robotic controller. In the simple example of a runner whose knee is blocked by a splint, the model predicts that serotonin specifically suppresses knee muscles, contradicting the generally accepted idea that serotonin affects all limb muscles equally. To test this prediction, human subjects performed fast and strong motions under the precise guidance of a robotic device. The resulting serotonergic effect was quantified and confirmed that serotonin scales the forces of individual muscles to maximize the motion amplitude. The presented results provide roboticists with a modular controller to boost the energy efficiency of cutting-edge elastic robots. For neuroscientists, the new understanding of serotonergic effects may enhance the rehabilitation of paraplegics. While state-of-theart therapies substitute serotonin in diffuse ways, the discovered precise effects can be mimicked by electric stimulation and exoskeletons to speed up the gait of patients. These far-reaching implications reveal the large potential of reverse-bionics to predict and explain new neuronal mechanisms. The numerous controllers of biomimetic robots can thereby reduce the frequently raised problem that neuroscience is data rich but theory poor.