New AI controller lets prosthetic legs adapt to any terrain without manual tuning

Powered prostheses have traditionally relied on impedance controllers that require extensive manual tuning and explicit mode classification. In a new preprint, researchers from Georgia Tech (Shim et al.) present a real-time end-to-end controller that directly estimates continuous actuator signals from onboard sensors—eliminating both tuning and intent classifiers. They trained Temporal Convolutional Networks on a multi-terrain dataset from 18 individuals with transfemoral amputation, then deployed the resulting model in real time across level ground, ramp ascent/descent, and stair ascent/descent.

Results from four participants (three able-bodied, one with amputation) showed the controller accurately reproduced key biomechanical relationships. During level walking, peak ankle torque scaled with speed (deployed 0.85 Nm/kg per m/s, p=0.001). On ramps, knee pre-flexion scaled with grade during ascent, and resistive knee torque increased during descent. For stairs, the controller generated seamless transitions even on sequences not previously seen in training. The work provides initial evidence that end-to-end neural control can deliver unified, mode-adaptive prosthetic assistance without subject-specific tuning—a step toward more natural and accessible bionic limbs.