Abstract
In the context of assistive robotics, myocontrol is one of the so-far unsolved problems of upper-limb prosthetics. It consists of swiftly, naturally, and reliably converting biosignals, non-invasively gathered from an upper-limb disabled subject, into control commands for an appropriate self-powered prosthetic device. Despite decades of research, traditional surface electromyography cannot yet detect the subject's intent to an acceptable degree of reliability, that is, enforce an action exactly when the subject wants it to be enforced. In this paper, we tackle one such kind of mismatch between the subject's intent and the response by the myocontrol system, and show that formal verification can indeed be used to mitigate it. Eighteen intact subjects were engaged in two target achievement control tests in which a standard myocontrol system was compared to two "repaired" ones, one based on a non-formal technique, thus enforcing no guarantee of safety, and the other using the satisfiability modulo theories (SMT) technology to rigorously enforce the desired property. The experimental results indicate that both repaired systems exhibit better reliability than the non-repaired one. The SMT-based system causes only a modest increase in the required computational resources with respect to the non-formal technique; as opposed to this, the non-formal technique can be easily implemented in existing myocontrol systems, potentially increasing their reliability.
Introduction
The "holy grail" of upper-limb prosthetics is myocontrol: using surface electromyography (sEMG) sensors on the skin to translate muscle signals into intuitive movements. Modern systems use Machine Learning (ML) to map these complex electrical patterns to control commands. However, ML models often suffer from intent mismatch. Imagine a user holding a glass; if it starts to slip, their natural instinct is to squeeze harder. In many ML models, this high-intensity signal falls outside the "training zone," causing the model's output to unexpectedly dip. The result is that the user squeezes harder, but the prosthetic relaxes its grip and drops the glass. This paper addresses the challenge of making these black-box ML models behave predictably and safely in high-force scenarios.
Methods
The authors propose a "repair" loop for ML models using formal verification. They start with a standard Ridge Regression model with Random Fourier Features. A safety property is defined: if the user's muscle activation ($t_A$) exceeds the maximum training value, the predicted output must remain $\ge 1$ (maximum force).
Using dReal, an SMT solver for non-linear functions, the system searches the model's input space for "unsafe points" where the model violates this property. If an unsafe point is found, synthetic data is generated for that point, and the model is retrained. This iterative process continues until the solver proves no more violations exist.
The SMT-repaired system was compared against an unrepaired baseline and a PDO (Physiologically Driven Optimization) system, which uses a simple heuristic search instead of rigorous formal proof to find errors.
Experimental Results
The study involved 18 intact subjects using a Thalmic Labs Myo bracelet. In a "stress test" designed to force users into the "danger zone" of high activation, the repaired models showed significant improvements. The success rate jumped from 14.58% (unrepaired) to approximately 46% with the repaired models. Furthermore, the "Reaching Rate" (RR)—measuring how often the user could hit the required activation level—improved from ~12% to over 42-46%.
The SMT repair process added roughly 15.32 seconds to the initial calibration phase, a negligible one-time cost for a prosthetic device, with zero latency during actual use.
Discussion and Conclusion
The core innovation is treating a non-linear ML model as a set of arithmetic constraints that can be formally verified. This moves myocontrol from being "statistically likely to work" to being "mathematically guaranteed to behave." While the heuristic PDO method performed similarly in these tests, the SMT approach provides a rigorous mathematical guarantee of safety crucial for medical devices.
Future work includes testing with amputee subjects, who may have noisier signals, and scaling the SMT solver to handle more complex "Degrees of Freedom," such as simultaneous finger and wrist movements. The authors also suggest using SMT to tackle action interference, preventing unintended movements during specific tasks.
Summary updated