AI-Powered Biomechanics: Revolutionizing Assistive Technologies

By Aryaman PattnayakReviewed by Susha Cheriyedath, M.Sc.Dec 19 2023

In an article published in the journal PLOS One, researchers propose that artificial neural networks (ANNs) can accurately and efficiently approximate solutions for the highly complex inverse dynamics computations required to represent realistic human arm and hand movements. The study highlights the computational demands in precisely calculating the biomechanics of the high degree-of-freedom human arm and hand system, emphasizing its essential role in advancing next-generation assistive technologies.

These technologies include advanced prosthetic devices, exoskeletons, and functional electrical stimulation systems. The researchers advocate for developing machine learning architectures to essentially serve as "artificial physics engines" for real-time simulation of limb kinetics. This proposed methodology could offer a more detailed, customizable, and reactive approach over traditional control techniques, facilitating more natural movement in impaired individuals.

The Study

Using detailed biomechanical modeling, the researchers first generated an extensive dataset representing the kinetics and kinematics of 23-degree-of-freedom upper extremity reaching movements. This included quantifying joint-space torques and angles over time from simulations of the arm traversing a 3D physiological workspace.

Various forms of recurrent neural networks - designed expressly for sequence processing tasks - were then trained on input vectors of just the kinematic trajectories to output predictions of the corresponding kinetic variables. The specific architectures analyzed included long short-term memory (LSTM) cells and gated recurrent units (GRUs). These have demonstrated excellent capabilities in complex statistical forecasting problems related to time-series data. 

After hyperparameter tuning, which involved adjusting factors such as learning rate and layer sizing, multiple design variations were tested. Among these, a 1-layer GRU network with 115 hidden nodes emerged as the optimal choice, yielding the lowest root mean square error (RMSE) in predicted joint torques against the actual reference values on the test dataset (RMSE under 0.03 Nm). The comparable single-layer LSTM architecture with the same hidden layer width also performed admirably, significantly outpacing the basic Elman RNN in accuracy.

Additional analysis around sequence length effects revealed that accuracy continued improving with more extended input history provided to the networks, finally saturating at under 0.1 Nm RMSE for GRU and LSTM with ≥50 previous timestep samples. So, while leverage of temporal information is beneficial, the networks can still function effectively with just fractions of a second of kinematics. Furthermore, testing revealed strong resilience to input noise, with only a minor 1% kinematic variability, increasing torque errors around 0.05 Nm for the GRU.

Together, these findings demonstrate that even simple recurrent architectures can successfully learn to solve exceedingly complex biomechanics problems when properly structured to leverage the temporal trajectory information inherent to coordinated movement. This supports the use of sequence-based neural network approaches as the foundation for data-driven modeling of limb dynamics. 

Advantages of Previous Work

Whereas most prior efforts at applying ANNs to approximate inverse dynamics for control have focused on relatively low dimensionality robots or basic joint systems with only ~5-7 degrees of freedom, this work pushes the boundary significantly further by tackling the far more complex coordinative dynamics of a fully articulated human arm and hand.

The proposed methodology could enable ANN-based control of highly sophisticated wearable assistive devices that match human movement capabilities and provide muscle-level reactive modulations closer to natural physiology. This is a crucial advancement, as the state-of-the-art systems today still struggle with delicate motor tasks and responding fluidly to rapid environment changes or stumbles. By essentially learning a surrogate of intricate muscle actuator coordination tailored to the individual and task, machine learning-based biomimetic controllers could help resolve this limitation.  

Future Outlook

While these initial results are auspicious, the investigation only probed a limited hyperparameter search space, so expanded analyses to automate ANN architecture tuning could uncover even better custom network designs. Additionally, training on a more diverse range of modeled behaviors beyond just simple reaching trajectories to skills like object manipulation could enhance model generalizability.

Input noise testing also surfaced substantially higher resilience for the GRU over LSTM models, so further work is warranted to determine if that advantage persists across task varieties. Finally, essential follow-on efforts could combine these latest ANN dynamics approximations with the researchers' previous human movement prediction models to provide a fully integrated end-to-end machine learning motor control system. 

In conclusion, this research demonstrates that recurrent neural networks can successfully serve as real-time "artificial physics engines" for simulating the tremendous complexity of human biomechanics with high accuracy. If complemented with detailed musculotendon modeling, the introduced methodology could be instrumental in developing the next evolution of advanced assistive systems that move and respond more naturally. This could dramatically improve the quality of life for many patients by restoring their capability to perform daily living activities.