Since IEC enables the user to continuously and directly modulate the prosthesis state, its inclusion in a volitional control architecture appears obvious. A general detriment to IEC is its constant reliance on user input. One can assume this increases cognitive user load compared to autonomously controlled devices, but this has yet to be quantified (Fleming et al., 2021). This potential downside applies to both IEC variants presented in the following.

The lack of immediate user authority over the device offered by CIC seems to rule this control approach out for volitional control at first glance. We still discuss CIC variants to get a comprehensive view on existing control approaches and evaluate their suitability to maybe be partially incorporated in a volitional approach.

Some studies combine IEC and CIC elements in hybrid prosthesis control strategies. For stair ascent with a knee prosthesis, Hoover et al. (2013) modulated the stiffness and equilibrium point of an impedance controller via both EMG input and predefined values for stance and swing phases. Wang et al. (2013) used EMG to manipulate plantarflexion force during push-off with an otherwise intrinsic controller in an ankle prosthesis, which enabled level-ground walking. This was adapted to also support stair ambulation by Kannape and Herr (2014). Shu et al. (2022) applied a similar control approach with an offline optimization method to map EMG inputs to the virtual muscles.

Hybrid control approaches allow for some volitional adjustments to the assistance given by the device. Because they do not solely rely on volitional input, they feature a potential reduction in cognitive load compared to full IEC. On the contrary, the presented methods still carry the disadvantages of finite state controllers. They rely on the detection of gait phases and are limited to predetermined movements and activities.

Human motor control signals are believed to consist of feed-forward and feedback components (Kandel et al., 2012). Markowitz et al. (2011) recognize this circumstance, but do not incorporate any user-controllable feed-forward signals in their control approach. Likewise, Shu et al. (2022) call reflexes essential for physiological gait and lament its absence in amputation musculature, but their device control relies on feed-forward signals alone. To our knowledge, the combination of muscle reflexes with feed-forward signals has only been tested in gait simulation, where it improved the model's robustness against perturbations compared to pure reflex control (Haeufle et al., 2018).

To facilitate compliant device behavior while limiting the cognitive demand on the user, we propose the development of a hybrid controller that combines EMG-based volitional input and simulated reflexes for the virtual muscle excitation in a model-based prosthesis control architecture. The structure of the proposed control approach is shown in Figure 2. Via EMG readings from residual muscles, the simulated muscles can be volitionally driven by the user, which subsequently command a desired torque to the prosthesis. Compared to the healthy limb, the afferent signals reporting muscle length (muscle spindles) and force (Golgi tendon organs) as well as cutaneous sensing to the spinal cord and brain (Kandel et al., 2012) are missing. This is substituted by incorporating virtual muscle states as well as sensory signals from the prosthesis hardware into simulated reflexes, which leads to an additional, non-volitional activation of the virtual muscles.

While this combined approach entails a high parameter count, it also allows for extensive individualization. Similar to the simulations conducted by Haeufle et al. (2018), the emphasis could be shifted between reflexes and volitional input by varying signal gains, depending on the wearer and their abilities to cognitively and physically handle the input responsibility. It is important to note that the studies mentioned here and in Section 2.2.2 mostly simulate monosynaptic stretch reflexes with a positive feedback of muscle force. This approach is self-reinforcing and would quickly lead to a maximum contraction of all muscles in a volitionally controllable prosthesis. Therefore, the reflex model needs to be adapted.

The lack of haptic and proprioceptive sensing in prostheses limits the natural and intuitive movement dexterity of users. Supplementary feedback can improve immediate control performance and promote the learning of internal motor models (Sensinger and Dosen, 2020). Feedback of the device state to the user is suggested to facilitate embodiment, device acceptance and dexterity (Beckerle et al., 2019).

Providing position feedback in an EMG-controlled position tracking task can heavily improve performance. Canino and Fite (2016) demonstrated this in a knee prosthesis with both, static pressure and vibratory feedback. The presented studies examining prosthetic gait did not employ such feedback mechanisms, while some mention the necessity of visual feedback. A completely different approach to proprioception restoration is the agonist-antagonist myoneural interface, which aims to recreate natural co-dependencies by attaching residual muscles to their respective antagonist in the residual limb (Clites et al., 2018). In addition to improved volitional muscle activation, this also enables the introduction of perceived passive movement by electrically stimulating the muscles. This surgical construct has enabled great control performance for individuals with an amputation, including the presented hybrid approach by Shu et al. (2022).

Given the discovered benefits of prosthesis state feedback to the user, it should receive more attention in lower-limb prosthesis research. Since advanced and novel surgery procedures will likely not be available to or suitable for everyone, feedback mechanisms built into devices need to be investigated and developed further. Current solutions for artificial feedback do not match natural fidelity, while selecting the appropriate feedback signals remains a challenge (Seminara et al., 2023). As pictured by Beckerle et al. (2017), the form and extent of the provided artificial feedback should be thoroughly assessed to ensure actual improvements in a given task. To enable conscious reactions to changes in the artificial limb state, the structure given in Figure 2 contains a feedback device, which receives the same simulated and sensor feedback as the artificial reflexes. It is ceded to future research to design the feedback derived from these signals.

Finding control parameters is a key challenge and choosing universal parameters for multiple users often does not appear suitable. The assistance required for a certain objective varies between individual users (Koller et al., 2015). People respond differently to active assistance and small changes can have large effects on energy expenditure (Voloshina and Collins, 2020). Thatte et al. (2018) observed that individual users prefer different gait characteristics and control parameter sets and called for methods to individualize prosthesis control.

Human-in-the-loop optimization (HILO) is a procedure of varying system parameters during user experiments to optimize for a given objective. It has been shown to successfully improve performance for exoskeletons and prostheses in a variety of tasks (Díaz et al., 2022) and offers the possibility to individualize prosthesis control (Voloshina and Collins, 2020). A common goal is the reduction of metabolic cost, which can be estimated from respiratory data (Zhang et al., 2017). An example for kinematic objectives is gait symmetry (Wen et al., 2020), which is especially interesting in the case of unilateral amputations. While HILO has been tested on a variety of control approaches for assistive devices, the combination with volitional control in lower-limb prostheses has yet to be realized.

We propose the use of HILO to determine control parameters that suit the individual needs of prosthesis users. Besides helping to manage large parameter spaces, HILO allows reseachers to pick different optimization objectives for different users or usecases. This can include desired gait kinetics, gait symmetry, metabolic cost, or user comfort and preference. For the last two, manual user input can be incorporated (Ingraham et al., 2020). Likewise, HILO allows to optimize for different tasks, like walking, stair-ascent and -descent or standing balance, depending on what is considered most important or most challenging by the user. To find initial parameters, anthropometric features and isometric torque and EMG measurements (Durandau et al., 2019) or predictive simulations of the human-machine interaction (Koelewijn and Selinger, 2022) can be used.

The studies on model-based EMG control mentioned in Section 2.1.2 all base their control on two simulated muscles in one degree of freedom (DoF) prostheses (plantar-/dorsiflexion for the ankle, flexion/extension for the knee). More elaborate models find use in CIC approaches (Markowitz et al., 2011; Thatte et al., 2018) and are being investigated for volitional control (Cimolato et al., 2020), but are not yet employed. Considering the example of an ankle prosthesis, active inversion/eversion was shown to reduce metabolic (Kim and Collins, 2017) and the lack of this DoF limits the adaption to uneven grounds. The use of only one plantarflexor muscle neglects the influence of the knee angle on the gastrocnemius length and hence the joint torque.

Challenging this common level of modeling depth could reveal potential benefits of more complex musculoskeletal models, though increase computational demands and parameter count. Therefore, the benefit of modeling certain biological anatomical features should be assessed and evaluated against the associated costs.