Stepping forward: Exploring biological control concepts to advance robotic locomotion

Kurzfassung

In biology, motions arise from the tight and complementary integration of morphology, control, and environment rather than hierarchical structures controlling individual joints. Understanding each aspect of this biological concept of embodied intelligence is a grand challenge that holds the promise of achieving equally elegant movements in robots. In this quest, sharing knowledge between these two fields can benefit both sides, as demonstrated in this dissertation for the control of dynamic compliant movements. Initial human experiments carried out as part of this dissertation revealed that the Central Nervous System (CNS) appears to apply topographically precise gain scaling of motoneurons, which functions analogously to a robotic autoencoder derived in previous research. This controller has been shown to be effective in exciting, simple intrinsic motions in compliant robots, suggesting that the CNS might do the same. This aligns with the idea that humans exploit their intrinsic body dynamics in motions, which was further supported by another human study presented in this work, showing that humans can intuitively estimate nonlinear dynamics and excite those in compliant interactions. To compare the suggested bio-plausible robotic controller with other bio-inspired controllers like Central Pattern Generators, both were implemented on a biomimetic leg. By comparing quantifiable metrics for energy efficiency and stability, this work highlighted that not only is the correct entrainment of the control signal to the mechanical system important, but the signal shape also needs to match the actuator. This inspired two further experimental investigations into modeling and exploiting motor dynamics and improving sim-to-real transfer for realistic robotic simulations. The frameworks from these investigations were combined with the bio-plausible robotic controller and recent advances in nonlinear mode theory to explore locomotion based on embodied intelligence in the novel highly-compliant quadruped eBert. Its mechanics encode six different modal oscillations, which can be validated on hardware using the bio-inspired simple robotic controller. Each modal oscillation suggested a matching foot pattern that was added with a black-box optimized step length. Through this, each mode naturally developed into a different gait, where the gait velocity related to the underlying mode frequency. Combined, this dissertation advanced biological understanding through robotic tools while applying new concepts for compliant robots inspired by biology, highlighting the potential of bi-directional knowledge transfer between the two fields.