A research proposal for an accessible, private, and accurate robotic prosthetic that returns user functionality to 100%, and continuously negotiates control, intent, and trust with its user.
Current intelligent prosthetics are expensive, inaccurate, and erode user privacy. This proposal designs a better approach, one that treats the prosthetic not as a tool, but as a co-robot.
This project began as an evaluation of the current state of intelligent prosthetics research — identifying the gaps that prevent existing devices from delivering on their promise. What emerged was a design proposal built around three core constraints: accessibility, privacy, and accuracy.
The central insight is that EMG signals, electrical signals produced by muscles, are noisy, user-specific, and context-dependent. They're hard to parse, but that difficulty is exactly what makes them interesting. Intelligence that works with messy, embodied signals is intelligence that works in the real world.
Four gaps that existing prosthetics fail to close.
Advanced prosthetics remain out of reach for most people who need them. The design prioritizes accessible hardware and edge computing so the intelligence lives on the device, not in an expensive cloud subscription.
Current devices achieve a high of 96% accuracy but require frequent recalibration as muscle signal patterns shift with fatigue, temperature, and time. This proposal trains for combined movements to build robustness instead of just discrete gestures.
EMG data is deeply personal, it encodes intent, emotion, and physical state. Current devices send this data to external servers. This proposal uses homomorphic encryption to enable personalization while keeping data on-device and private.
Most prosthetics support a limited set of discrete gestures. Real human movement is continuous, combined, and context-dependent. The goal of 100% functionality return requires a fundamentally different model of how the device interprets intent.
"This solution will be a co-robot, that continuously negotiates control, intent, and trust with its user. It is work on the body as assistive technology."
A cascaded model approach: a generic model trained on a broad population provides the baseline, while a personal model fine-tunes on the individual user's signal patterns, adapting over time without requiring full recalibration.
Four specific ideas that address the gaps in current research.
Train for combined movements, not discrete gestures. Instead of classifying pinch, up, down as separate states, train the model on the continuous space of combined movements, closer to how the human nervous system actually operates.
Gather data from both limbless patients and a control group to build a robust training dataset that generalizes across signal variation, not just within a single population. This directly addresses the recalibration brittleness in current devices.
Cascade ML models a generic model trained on the full population, with a personal model layered on top that adapts to the individual user's signal patterns over time. Personalization without full recalibration.
Enable real-time data feedback with personal recalibration offloaded to a nearby edge device such as a smartphone, therefore keeping latency low, keeping data private, and making advanced ML inference accessible without cloud dependency.
"EMG is noisy, user-specific, and context-dependent. It confronts how intelligence emerges from difficult-to-parse, embodied signals and that is exactly the problem worth solving."
This is a co-robot, not a tool. The framing matters. A tool executes commands. A co-robot negotiates — it infers intent, asks for confirmation, adapts over time. That shift in framing changes every design decision downstream.
Agency is the central design question for personal robotics. Where does human intent end and robotic autonomy begin? Designing an EMG-controlled limb forces you to answer that question concretely, in real time, for a real person.
Privacy is not a constraint, it is a feature. Homomorphic encryption enables the device to learn and personalize without ever exposing raw EMG data. Privacy-preserving personalization is the standard this field should be building toward.
The intersection of distributed computing and ML is where this becomes real. The pipeline I built at Microsoft and the ML work at Kwartile converge here — real-time data infrastructure, at the edge, for a human body.