Neuromorphic control of an inverted pendulum

Details

Supervisors

Bianchin, Gianluca ; Delvenne, Jean-Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

The brain stabilises a standing body on about twenty watts, while modern robots spend kilowatts to do less, and with a fraction of the robustness. This energetic gap is not an artefact of immature engineering but a structural one, and it points to a different paradigm of control: one that is rhythmic and event-driven rather than continuously sampled, and adaptive by construction. This thesis asks whether such a paradigm, transcribed from three well-identified biological mechanisms (central pattern generation, neuromodulatory regulation, and sensorimotor phase correction), can stabilise the canonical benchmark of balancing tasks, the inverted pendulum. The work pursues two coupled objectives. The first is a systematic bifurcation analysis of the half-centre oscillator, a pair of reciprocally inhibiting neurons producing an antiphase rhythm, in the slow-/ultra-slow-gain plane; it identifies a single bursting zone as the operating window of any event-based controller built on this primitive. The second uses this design map to construct an event-based neuromorphic controller that stabilises the inverted pendulum at prescribed amplitude and frequency. The inverted pendulum is the elementary abstraction of a biped in single-leg stance. The resulting closed loop absorbs velocity shocks, friction noise, torque noise and nearly five decades of sensing delay, while spending mechanical effort comparable to an LQR baseline and requiring roughly one third of the host-side operations. Together, these results trace an architectural path from a toy unstable plant to a neuromorphic walking machine, and recast the engineering transcription of biological control mechanisms as a methodology rather than a metaphor, one equipped with its own design map and operating window. Taken further, the paradigm points toward embedded controllers in which adaptation, robustness and energetic frugality no longer compete, and toward autonomous machines that begin to close the orders-of-magnitude gap nature still holds over engineering, from locomotion and balance to respiration, manipulation, and sensorimotor coordination.