For applications requiring interaction with humans or unstructured environments, robots are increasingly designed to leverage the intentional drivetrain compliance of series elastic actuators (SEAs). Impedance control, likewise, is of particular value in these applications, having long been considered an effective means of addressing dynamic interaction. Although impedance controlled SEAs appear often in the literature, a number of important questions remain unanswered. If, for example, robust contact stability is required, can an SEA render a virtual stiffness greater than its physical spring rate? Previous studies answer no, but this is largely a question of control architecture. It is proven here, as part of a larger study comparing the stability and passivity of five different control approaches, that this is in fact possible if disturbance observer based impedance control is adopted.

The fidelity with which SEAs render desired impedances is important as well. In comparing the impedance rendering accuracy of multiple control approaches, experimental data in both the time and frequency domain point once more to disturbance observer based impedance control. This new SEA control architecture yields demonstrable improvement in actuator transparency, closed loop hysteresis, and the SEA's dynamic response to both reference commands and external torques.

Two new performance metrics are formulated based on the H∞ and H2 system norms to further quantify SEA impedance rendering accuracy across the frequency spectrum. A novel model matching framework is then constructed that leverages these metrics for the optimal synthesis of SEA impedance control. Passive and accurate controllers result that, having been deployed on physical hardware, represent the first application of LMI-based, multi-objective, optimal control synthesis to series elastic actuation.

These results are all confirmed experimentally on high performance SEAs. Three new actuator designs are presented that provide up to 350 Nm in peak torque and torque sensing resolutions as low as 0.006 Nm. This 58,333:1 dynamic range (an order of magnitude improvement over previous SEAs) is achieved in a torque dense, 94.3 Nm/kg package ideally suited for use in humanoid robots. Demonstrated SEA performance reinforces the practical utility of the recommended control approaches and speaks to the broader applicability of impedance controlled SEAs to human-centric robots.