Tesi etd-07092024-181446
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
EKEN, HÜSEYIN
URN
etd-07092024-181446
Titolo
Control of Lower-Limb Prostheses and Exoskeletons based on adaptive Dynamic Movement Primitives
Settore scientifico disciplinare
ING-IND/34
Corso di studi
Istituto di Biorobotica - PHD IN BIOROBOTICA
Commissione
Presidente Prof. VITIELLO, NICOLA
Membro Prof.ssa CASADIO, MAURA
Membro Prof.ssa CASADIO, MAURA
Parole chiave
- wearable robotic devices
- locomotion mode recognition
- continuous gait phase estimation
Data inizio appello
12/02/2025;
Disponibilità
parziale
Riassunto analitico
As the primary source of mobility and self-reliance in daily life, impairments in lower limb function can profoundly impact individuals, leading to short- and long-term physiological, psychological, and socio-economic problems. Recent technological advancements, particularly in wearable robotic devices, offer promising solutions to support individuals suffering from lower limb loss and dysfunction in everyday activities and rehabilitation processes. The effectiveness of lower-limb wearable robots hinges on two main factors: their mechanical characteristics (such as actuator performance, weight, and size) and the underlying control strategies. Current devices integrate intelligent mechanical systems and capable actuation strategies. On the control side, several strategies have proven effective in interfacing with the human body, translating user intentions or biomechanical data into coordinated movement patterns aligned with user needs and intended motions.
To advance the current state-of-the-art, this thesis focuses on the design, development, and validation of control algorithms, particularly emphasizing the adaptive Dynamic Movement Primitives (aDMP) algorithm. aDMP are error-driven dynamic systems capable of encoding general movement behaviors and continuously estimating task features such as trajectory, target goal (trajectory value at the end of the movement), and movement phase in real time. Initially developed for discrete and monotonic movements, aDMP had limited applicability to real-life lower limb movements. This thesis extended the applicability of aDMP to encompass both discrete and rhythmic, and monotonic and non-monotonic movement profiles, thus, significantly broadening their utility. The generalized aDMP were then applied to recognize the locomotion mode of individuals and continuously estimate the movement phase for both rhythmic and discrete movements.
In the locomotion mode recognition, aDMP were employed as locomotion templates for feature extraction for a group of Support Vector Machines (SVMs). The algorithm was tested in an offline analysis with data from ten able-bodied participants and a subject with transtibial amputation to recognize stair ascending, level-ground walking, and stair descending locomotion. Moreover, when classifying transition steps, the algorithm leveraged the biomechanical similarities between the steady-state locomotion. Results showed that the algorithm could recognize the investigated modes when even the locomotion was first initiated from a standing position, at no later than the first half of the swing phase with high accuracies (up to 100%). Nevertheless, it struggled to accurately recognize transition steps (median accuracies up to 87.30% and 93.10% for the able-bodied participants and the subject with transtibial amputation, respectively), which also occurred at the late stages of the swing phase. Furthermore, benchmarking the proposed approach against neural networks revealed that incorporating aDMP effectively reduced the complexity of the classification algorithm: simple SVMs achieved comparable accuracy to fully connected and deep neural networks, while coupling aDMP with these networks enhanced the classification of transitional steps.
In continuous phase estimation, the efficacy of aDMP was first investigated across multiple locomotion activities (including level-ground walking, stair negotiation, and ramp negotiation), involving five able-bodied participants wearing a sensory system composed of pressure-sensitive insoles and inertial measurement units. The estimated phase resulted in a median root-mean-square error of 3.34 [3.19, 4.53] %GC and a final estimation error of 1.00 [0.00, 1.00] %GC, with respect to the linear phase, across all locomotion activities. Notably, the gait phase was also estimated accurately during nearly all transitions, though performance indicators for these steps were slightly elevated relative to overall performance. These results provided a viable foundation for future phase-dependent control strategies for lower-limb wearable robots. Moreover, in a manner analogous to the locomotion mode recognition study, the performance of the aDMP algorithm was benchmarked offline against five prominent neural network architectures from the state-of-the-art. While no statistically significant differences were observed between aDMP and three of these networks, two of the neural networks outperformed the aDMP approach. Given the offline nature of this investigation and the relatively small sample size, further real-time studies with larger populations are needed to fully assess the comparative performance of aDMP against these neural network methods.
Further advancing the application of aDMP, a subsequent study explored generating phase-dependent torque profiles during level-ground walking with a unilateral hip exoskeleton for six able-bodied individuals. This study assessed the performance of the aDMP algorithm using subject-dependent and subject-independent models. In parallel, the study assessed the offline performance of aDMP for estimating gait phase in pathological gait using data from six stroke survivors, and for discrete tasks—specifically sit-to-stand and stand-to-sit transitions—in five able-bodied individuals. Results showed root-mean-square errors that ranged from 2.75%GC to 5%GC relative to the linear phase across all experimental conditions (including assistive and transparent modes with sub-ject-dependent and subject-independent models) for level-ground walking of able-bodied participants and a median of 6.75%GC for stroke survivors. For sit-to-stand and stand-to-sit transitions, subject-independent models yielded root-mean-squared error values exceeding 10% of the movement duration; however, a semi-subject-dependent approach, which tailored movement primitives to each subject while using common model parameters, reduced this error by half to below 5% of the movement duration. These findings suggest that aDMP-based gait phase estimation, particularly with the subject-independent model, can effectively generate phase-dependent torque profiles for able-bodied participants, however, further research is needed to validate these findings in stroke survivors due to the higher errors observed in this population. Moreover, the promising results in discrete tasks highlight the potential of the semi-subject-dependent approach, paving the way for aDMP utilization in multi-modal clinical assessments such as the Timed Up and Go test.
To solidify these findings of the previous study, a third study was conducted with 12 able-bodied participants who performed similar level-ground walking tasks with the same unilateral hip exoskeleton. The results indicated that the subject-dependent and subject-independent aDMP models produced statistically similar performances in gait phase estimation and assistance profile generation, thereby validating the earlier findings with a larger participant cohort. Specifically, median root-mean-square errors were 2.43%GC and 3.90%GC for the subject-dependent model and 3.05%GC and 3.01%GC for the subject-independent model in transparent and assistive mode trials, respectively. The subject-independent model demonstrated notable resilience to variations in input angles caused by the generated assistance, making them more preferable compared to subject-dependent models.
This study then compared the performance of the subject-independent aDMP model with Adaptive Oscillators (AOs), a widely used method known for producing linear and smooth phase estimates during steady-state gait. The conducted analytical comparison confirmed the theoretical strengths and weaknesses of both approaches: aDMP provided mostly consistent performance throughout the movement, albeit with less line-arity and smoothness compared to AOs, while AOs struggled during the initial movement stages (e.g., median root-mean-square errors of 11.16%GC compared to 3.16%GC of the aDMP model in transparent mode trials) but delivered more accurate phase estimates during the later steady-state stages of movement (e.g., median root-mean-square errors of 1.11%GC compared to 3.09%GC of the aDMP model in the same trials).
Finally, this study introduced a novel hybrid phase estimation algorithm that builds on these previous findings by combining aDMP and AOs to mitigate the drawbacks of each method while leveraging their strengths. The hybrid algorithm employs aDMP during the initial stages of movement, where AOs typically underperform. As the movement progresses and AOs become capable of providing more accurate phase estimations, the hybrid algorithm seamlessly transitions from aDMP to AOs. This approach ensures that the proposed hybrid algorithm delivers phase-dependent assistance profiles right from the onset of movement, outperforming the use of AOs alone, and provides more accurate and linear phase estimates during prolonged use compared to using aDMP alone. Indeed, the proposed hybrid algorithm resulted in a median root-mean-square error of 1.07 %GC throughout the movement, which significantly overperformed both the aDMP and AOs algorithms.
In conclusion, this thesis work underscores the extended application of the aDMP model across various lower-limb movement types, highlighting its potential to enhance the functionality of wearable robotic devices in two distinct control scenarios. These demonstrated scenarios are crucial in enabling multi-modal functionality and providing control strategies that are robust to movement speed variations and tailored to different movement phases in wearable robotics. The versatility of the aDMP model is highlighted by the distinct nature of these control scenarios. The studies presented in this thesis work demonstrate comparable performance with respect to current state-of-the-art alternatives, paving the way for further research to apply the aDMP algorithm in pathological end-users and to explore its integration across both control scenarios.
To advance the current state-of-the-art, this thesis focuses on the design, development, and validation of control algorithms, particularly emphasizing the adaptive Dynamic Movement Primitives (aDMP) algorithm. aDMP are error-driven dynamic systems capable of encoding general movement behaviors and continuously estimating task features such as trajectory, target goal (trajectory value at the end of the movement), and movement phase in real time. Initially developed for discrete and monotonic movements, aDMP had limited applicability to real-life lower limb movements. This thesis extended the applicability of aDMP to encompass both discrete and rhythmic, and monotonic and non-monotonic movement profiles, thus, significantly broadening their utility. The generalized aDMP were then applied to recognize the locomotion mode of individuals and continuously estimate the movement phase for both rhythmic and discrete movements.
In the locomotion mode recognition, aDMP were employed as locomotion templates for feature extraction for a group of Support Vector Machines (SVMs). The algorithm was tested in an offline analysis with data from ten able-bodied participants and a subject with transtibial amputation to recognize stair ascending, level-ground walking, and stair descending locomotion. Moreover, when classifying transition steps, the algorithm leveraged the biomechanical similarities between the steady-state locomotion. Results showed that the algorithm could recognize the investigated modes when even the locomotion was first initiated from a standing position, at no later than the first half of the swing phase with high accuracies (up to 100%). Nevertheless, it struggled to accurately recognize transition steps (median accuracies up to 87.30% and 93.10% for the able-bodied participants and the subject with transtibial amputation, respectively), which also occurred at the late stages of the swing phase. Furthermore, benchmarking the proposed approach against neural networks revealed that incorporating aDMP effectively reduced the complexity of the classification algorithm: simple SVMs achieved comparable accuracy to fully connected and deep neural networks, while coupling aDMP with these networks enhanced the classification of transitional steps.
In continuous phase estimation, the efficacy of aDMP was first investigated across multiple locomotion activities (including level-ground walking, stair negotiation, and ramp negotiation), involving five able-bodied participants wearing a sensory system composed of pressure-sensitive insoles and inertial measurement units. The estimated phase resulted in a median root-mean-square error of 3.34 [3.19, 4.53] %GC and a final estimation error of 1.00 [0.00, 1.00] %GC, with respect to the linear phase, across all locomotion activities. Notably, the gait phase was also estimated accurately during nearly all transitions, though performance indicators for these steps were slightly elevated relative to overall performance. These results provided a viable foundation for future phase-dependent control strategies for lower-limb wearable robots. Moreover, in a manner analogous to the locomotion mode recognition study, the performance of the aDMP algorithm was benchmarked offline against five prominent neural network architectures from the state-of-the-art. While no statistically significant differences were observed between aDMP and three of these networks, two of the neural networks outperformed the aDMP approach. Given the offline nature of this investigation and the relatively small sample size, further real-time studies with larger populations are needed to fully assess the comparative performance of aDMP against these neural network methods.
Further advancing the application of aDMP, a subsequent study explored generating phase-dependent torque profiles during level-ground walking with a unilateral hip exoskeleton for six able-bodied individuals. This study assessed the performance of the aDMP algorithm using subject-dependent and subject-independent models. In parallel, the study assessed the offline performance of aDMP for estimating gait phase in pathological gait using data from six stroke survivors, and for discrete tasks—specifically sit-to-stand and stand-to-sit transitions—in five able-bodied individuals. Results showed root-mean-square errors that ranged from 2.75%GC to 5%GC relative to the linear phase across all experimental conditions (including assistive and transparent modes with sub-ject-dependent and subject-independent models) for level-ground walking of able-bodied participants and a median of 6.75%GC for stroke survivors. For sit-to-stand and stand-to-sit transitions, subject-independent models yielded root-mean-squared error values exceeding 10% of the movement duration; however, a semi-subject-dependent approach, which tailored movement primitives to each subject while using common model parameters, reduced this error by half to below 5% of the movement duration. These findings suggest that aDMP-based gait phase estimation, particularly with the subject-independent model, can effectively generate phase-dependent torque profiles for able-bodied participants, however, further research is needed to validate these findings in stroke survivors due to the higher errors observed in this population. Moreover, the promising results in discrete tasks highlight the potential of the semi-subject-dependent approach, paving the way for aDMP utilization in multi-modal clinical assessments such as the Timed Up and Go test.
To solidify these findings of the previous study, a third study was conducted with 12 able-bodied participants who performed similar level-ground walking tasks with the same unilateral hip exoskeleton. The results indicated that the subject-dependent and subject-independent aDMP models produced statistically similar performances in gait phase estimation and assistance profile generation, thereby validating the earlier findings with a larger participant cohort. Specifically, median root-mean-square errors were 2.43%GC and 3.90%GC for the subject-dependent model and 3.05%GC and 3.01%GC for the subject-independent model in transparent and assistive mode trials, respectively. The subject-independent model demonstrated notable resilience to variations in input angles caused by the generated assistance, making them more preferable compared to subject-dependent models.
This study then compared the performance of the subject-independent aDMP model with Adaptive Oscillators (AOs), a widely used method known for producing linear and smooth phase estimates during steady-state gait. The conducted analytical comparison confirmed the theoretical strengths and weaknesses of both approaches: aDMP provided mostly consistent performance throughout the movement, albeit with less line-arity and smoothness compared to AOs, while AOs struggled during the initial movement stages (e.g., median root-mean-square errors of 11.16%GC compared to 3.16%GC of the aDMP model in transparent mode trials) but delivered more accurate phase estimates during the later steady-state stages of movement (e.g., median root-mean-square errors of 1.11%GC compared to 3.09%GC of the aDMP model in the same trials).
Finally, this study introduced a novel hybrid phase estimation algorithm that builds on these previous findings by combining aDMP and AOs to mitigate the drawbacks of each method while leveraging their strengths. The hybrid algorithm employs aDMP during the initial stages of movement, where AOs typically underperform. As the movement progresses and AOs become capable of providing more accurate phase estimations, the hybrid algorithm seamlessly transitions from aDMP to AOs. This approach ensures that the proposed hybrid algorithm delivers phase-dependent assistance profiles right from the onset of movement, outperforming the use of AOs alone, and provides more accurate and linear phase estimates during prolonged use compared to using aDMP alone. Indeed, the proposed hybrid algorithm resulted in a median root-mean-square error of 1.07 %GC throughout the movement, which significantly overperformed both the aDMP and AOs algorithms.
In conclusion, this thesis work underscores the extended application of the aDMP model across various lower-limb movement types, highlighting its potential to enhance the functionality of wearable robotic devices in two distinct control scenarios. These demonstrated scenarios are crucial in enabling multi-modal functionality and providing control strategies that are robust to movement speed variations and tailored to different movement phases in wearable robotics. The versatility of the aDMP model is highlighted by the distinct nature of these control scenarios. The studies presented in this thesis work demonstrate comparable performance with respect to current state-of-the-art alternatives, paving the way for further research to apply the aDMP algorithm in pathological end-users and to explore its integration across both control scenarios.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|