Bipedal robots and prosthetics usually leave the toe out entirely, since a dedicated toe actuator adds weight and complexity. Industry robots like Unitree's G1 and Boston Dynamics' Electric Atlas show that quasi-direct-drive (QDD) actuation and modern learned controllers can produce a fairly stable gait even with flat, rigid feet, but that rigidity forces the sole to stay parallel to the ground through stance, removing heel-strike, pushing more load onto the knees, and costing walking efficiency and stride length/speed (one biomechanics study found robots with toes achieve roughly 13% higher energy efficiency and longer strides than toeless designs). Existing powered-toe designs address this either with a full second actuator (which in current prosthetic and full-leg designs tends to be heavy and mounted proximally on the leg rather than at the foot, ruling it out as a prosthetic) or with a passive/variable-stiffness toe that stays light and allows heel-strike but isn't actually powered and can't inject positive push-off force or active balance control. My thesis looks for a middle path: real, controllable toe push-off power without the weight and complexity cost of a second dedicated actuator, designed to be light and distal enough to work as a prosthetic, not just on a robot.

Since the toe really only needs power during the push-off portion of the gait cycle, I designed the mechanism around underactuation: prioritizing toe control specifically during push-off, and trading off some control during the rest of the gait cycle when the toe matters less. The one existing underactuated ankle-toe design in the literature (from the University of Utah) takes a similar tack but only recycles energy through a movable toe to reduce motor torque; it never injects positive power at the toe. The mechanism I'm developing is a novel ankle-toe linkage driven by a single QDD actuator, using a shrinking-bar linkage to modulate how the ankle actuator drives the toe, mechanically coupling the ankle and toe while still retaining control over how the two move relative to each other. The design goal is to reduce total actuator count for an ankle-toe system while still injecting real, positive toe push-off power and maintaining control over actuation timing and speed. That's the core research question: can a linked ankle-toe mechanism that draws on a single motor's power for both joints provide stable walking while still delivering the benefits of a powered toe?

I'm using reinforcement learning to explore and quantify the benefits of this design compared to traditional and alternative ankle-toe designs, evaluating things like walking energy efficiency and stride characteristics against a toeless baseline and other powered/passive toe approaches.