Supervised Learning Approaches for Static and Temporal Force Prediction in a Biohybrid Muscle Ring Actuator

summary

The gist

Skeletal muscle-based biohybrid actuators anchored on flexible polymer pillars were studied under static and dynamic machine learning prediction models.

In short

This study used machine learning to predict force from skeletal muscle-based biohybrid actuators anchored on polymer pillars. Static models predicted maximum force with high accuracy (R2 up to 0.9425). A dynamic LSTM model accurately reproduced the full time series of force response, achieving an R2 of 0.9956, showing ML's potential for modeling complex biological systems.

Key concepts

Static Prediction Models
These models estimate the maximum possible force a muscle ring can exert before stimulation by using inputs like muscle sample number, stimulation frequency, and waveform type. They help optimize actuator performance for specific force requirements.
Dynamic Modeling Framework (LSTM)
This framework uses a Long Short-Term Memory network to create a digital twin of the actuator. It learns to replicate the entire sequence of force measurements over time in response to electrical stimulation, which is crucial for understanding temporal behavior.
R2 Score
The R2 score measures how well the model's predictions match the actual experimental data. A high R2 score indicates that the machine learning model is highly accurate at predicting both static force limits and dynamic force changes.
Digital Twin
In this context, a digital twin is a virtual replica of the physical biohybrid muscle actuator. The LSTM model acts as this twin, allowing researchers to simulate and predict how the physical device will behave under different stimulation conditions.

Terminology used across episodes

This episode discusses

The paper

Supervised Learning Approaches for Static and Temporal Force Prediction in a Biohybrid Muscle Ring Actuator · Read on arXiv

Michail-Antisthenis Tsompanas, Marco Perez Hernandez, Faisal Abdul-Fattah, Karim Elhakim, Mostafa Ibrahim, Judith Fuentes, Florencia Lezcano, Riccardo Collu, Massimo Barbaro, Stefano Lai, Samuel Sanchez

Unconventional Computing Laboratory & School of Computing & Creative Technologies, University of the West of England, Bristol Institute for Bioengineering of Catalonia (IBEC), The Barcelona Institute for Science and Technology (BIST), Department of Electrical and Electronic Engineering, University of Cagliari, Institució Catalana de Recerca i Estudis Avancats (ICREA)

Skeletal muscle-based biohybrid actuators have proved to be a promising component in soft robotics, offering efficient movement. However, their intrinsic biological variability and nonlinearity pose significant challenges for controllability and predictability. To address these issues, this study investigates the application of supervised learning, a form of machine learning, to model and predict the behavior of biohybrid machines (BHMs), focusing on a muscle ring anchored on flexible polymer pillars. First, static prediction models (i.e., random forest and neural network regressors) are trained to estimate the maximum exerted force achieved from input variables such as muscle sample, electrical stimulation parameters, and baseline exerted force. Second, a dynamic modeling framework, based on Long Short-Term Memory networks, is developed to replicate the time series of exerted forces observed. Both modeling approaches demonstrate high predictive accuracy. The best performance of the static models is characterized by a mean R2 of 0.9346, whereas the dynamic model achieves a mean R2 of 0.9944 over five training instances. The static models can enable optimization of muscle actuator performance for targeted applications and required force outcomes, while the dynamic model provides a foundation for developing robustly adaptive control strategies in future biohybrid robotic systems.

DOI: 10.1002/aisy.70637

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Supervised Learning Approaches for Static and Temporal Force Prediction in a Biohybrid Muscle Ring Actuator".

Title and authors: Dev: The authors are listed as Michail Antisthenis Tsompanas, Marco Perez Hernandez, Faisal Abdul-Fattah, Karim Elhakim, Mostafa Ibrahim School of Computing and Creative Technologies at the University of the West of England in Bristol. It's a team from computing and engineering backgrounds.

Rosa: That’s right. They are bringing together computational methods with the biological hardware here. The title tells us exactly what they’re doing: using supervised learning for static and temporal force prediction in biohybrid muscle ring actuators, which is pretty specific.

Taro: What does that actually mean for a robot that might have some unpredictable movement? Does it just give us a better guess of the max force it can push?

Dev: It means they are trying to predict the maximum possible exerted force on the pillars based on things like the muscle sample, electrical stimulation parameters, and baseline behavior before they even apply any stimuli. That’s what that static prediction is for.

Rosa: Exactly. They found that their best static models got an R two score of zero point nine four two five, while the dynamic model achieved an R two score of zero point nine nine five six when predicting the full time series of exerted force in response to electrical stimulation, which is really high accuracy for a dynamic system <ref:2602.16330#pg1>.

Taro: A prediction that works well during steady-state operation but then drops off when there are sudden changes, you know, if something unexpected happens?

Dev: That’s the caveat they found. The results show that forecast accuracy is high and stable during steady-state operation but deteriorates at sudden changes, suggesting a need to incorporate dynamic uncertainty into the digital twin they are building.

The paper's summary: Rosa: So, putting it together, this paper summarizes how they used two complementary modeling strategies to handle the challenges of biological variability and nonlinearity in biohybrid actuation. They looked at static prediction using random forest and neural network models to estimate the maximum possible exerted force based on muscle sample, electrical stimulation parameters, and baseline behavior before stimuli application.

Dev: And then they introduced a dynamic modeling framework using a long short term memory network to construct a digital twin capable of reproducing the full time series of exerted force in response to electrical stimulation. They’re treating it as a univariate time series forecasting problem.

Taro: So, the static models give us an estimate for the maximum force, and the dynamic model gives us what happens over time when we actually stimulate it. How does that help someone who only listens to this show?

Rosa: For someone listening, it means we get a tool to model these muscle actuators. The static part helps optimize performance for targeted applications by estimating the maximum force per experiment based on inputs like the sample number of the muscle ring, frequency, pulse width and waveform type of stimulation signals.

Dev: And the dynamic model is what lets them reproduce the full time series exerted force, which means it’s a foundation for developing robustly adaptive control strategies in future ML-driven biohybrid robotic systems. That’s where the real control potential lies.

The paper's improvements: Taro: I saw some details about how they improved their static model, like adding that extra input feature of the average of the first ten time steps recorded for each experiment. What was the impact of that change?

Rosa: They found that incorporating baseline exerted force prior to stimulation further improved model accuracy by reducing the mean square error by nine point six percent relative to the initial neural network and by twenty point four percent compared to a fine tuned random forest model <ref:2602.16330#pg3>. That’s a solid improvement for handling those inter-sample offsets we talked about earlier, Taro.

Dev: It sounds like that baseline force input was key because they noted there was considerable variation in baseline force across experiments and muscle samples; by supplying this baseline explicitly, the network could better adjust for those offsets. That makes sense from a control perspective.

Taro: So it’s about making the static prediction more robust against those natural differences between different muscle samples or setups in the lab?

Rosa: Precisely. The residual analysis confirmed that all their static models showed a random and homogeneous error distribution, which suggests that while they improved accuracy, there isn't a hidden systematic bias left in their predictions for those maximum force estimates.

Conclusion: Dev: So to wrap up on this paper, the implication is that machine learning tools are expected to have a key role in future designing processes and optimization of biohybrid machines because deterministic modeling is difficult due to high biological variability and nonlinear dynamics.

Rosa: They demonstrated that ML models can effectively predict these outputs through both static and dynamic training on experimental data, with the dynamic prediction model reproducing the full time series of force achieved by the actuator with an r two score of zero point nine nine five six <ref:2602.16330#pg1>.

Taro: I think what this really changes for us is that we now have a way to use these models to guide how we design and optimize these systems, even if they aren't perfect in every scenario.

Dev: Yeah, the dynamic model provides a foundation for developing robustly adaptive control strategies in future biohybrid robotic systems. We can use that time series prediction as a baseline for whatever control loop we build on top of it.

Rosa: That’s the big picture. We learned how to use supervised learning to handle these complex biological dynamics, and we got high fidelity results for both static and dynamic force prediction in this paper, "Supervised Learning Approaches for Static and Temporal Force Prediction in a Biohybrid Muscle Ring Actuator."

More episodes

← Home