When robot systems enter into a mechanical feedback relationship with humans, the closed-loop system must remain stable despite any unmodeled high-frequency dynamics in order to be deemed safe, and this imposes a bandwidth limit on performance. However, humans are difficult to model accurately for feedback purposes. Their behavior is never perfectly known nor is it consistent. My team’s work at The University of Texas at Austin has shown that if a robot can infer the changing human compliance, it can update its control strategy to significantly improve strength amplification performance [HuangCappelThomasHeSentis2020ACC]. However, there is a fundamental gap in knowledge on how to reconcile the desire for formal stability guarantees with the adaptive control theory that allows feedback controllers to make these changes. I hypothesize that these goals can be simultaneously achieved by the online synthesis of both a bounded-uncertainty model to explain past measurements and a controller that robustly stabilizes this model. My preliminary work towards this topic \cite{ThomasSentis2019TAC} has demonstrated for the first time that direct synthesis of a bounded-uncertainty model is possible and that it is even a convex optimization problem, a problem class that can be solved reliably by numerical algorithms. Tools exist to design adaptive controllers that are robust in the sense of modeled disturbances of known relative order but not in the sense of guaranteed stability despite unmodeled dynamics. The overall goals of this research thrust are to (1) extend the framework for robust model identification into a complete theoretical framework for online synthesis of uncertain models and robust controllers, (2) establish the feasibility of adapting exoskeleton controllers to identified uncertain human models in a 2-DOF strength amplifying co-bot testbed, and (3) determine the extent to which the framework can scale to larger systems given realistic limits on computational power. This research is aligned with the mission of the NSF CMMI program in Dynamics, Control and Systems Diagnostics.
Research
Assistive exoskeletons for activities of daily life
Lightweight assistive exoskeletons have incredible potential to improve quality of life for those with mobility impairments such as sarcopenia and osteoarthritis [MedranoRouseThomas2021TMRB]; however, several challenges are hindering their translation to the home and workplace. One key limitation is that assistive exoskeleton control strategies are predicated on only operating in a small, fixed set of tasks (e.g., walking, stair ascent, etc.). This limitation prevents exoskeletons from handling the broad scope of locomotion modes—and the fluid transitions between them—that are essential for activities of daily living.
A task-invariant control framework is therefore needed to develop practical exoskeleton technology that can assist humans in their everyday lives. This framework needs to be able to recognize human intent without assuming a pattern in their behavior, and to address this point I proposed to use shoe-sole force/torque sensors in conjunction with inertial measurement units to estimate the human’s joint torques directly from the physics of the human body [ThomasGregg2021CDC]. In this research direction, my aims are to (1) establish a theoretical framework for the stable control and energetic passivity properties of exoskeletons using joint torque estimates inferred from ground reaction force sensor, (2) design and build a lightweight backdrivable hip-knee-ankle exoskeleton similar to the M-BLUE modular orthosis [NeslerThomasDivekarRouseGregg2021RAL] and a force sensing shoe, and (3) determine the extent to which wearers can perceive the amplification bandwidth of the exoskeleton (a potential performance benchmark for task-invariant controllers) using a psychophysical experiment. As it addresses a fundamental gap in robotics capabilities, this research aligns with the goals of the NSF CMMI program on Foundational Research in Robotics. Successful strength amplification would both reduce muscle tension and, depending on the mechanical design of the exoskeleton, apply loads to the joints in proportion to the applied torque [MedranoRouseThomas2021TMRB]. Both of these effects could be exploited to alter joint loading in beneficial ways, and their optimization is a potential collaboration opportunity with biomedical researchers interested in osteoarthritis.
Compliant Actuators for a Space Modular Manipulator
In collaboration with the RAD lab, we are developing a series of modular arm components that are space-rated and can be rapidly assembled into a large array of manipulators. A key feature of the project is series-elastic actuation, which will enable high fidelity torque control and multi-contact manipulation in micro-gravity. Our approach builds on previous work with the control of NASA Valkyrie actuators [1].
[1] Thomas GC, Mehling JS, Holley J, Sentis L. Phase-relaxed-passive full state feedback gain limits for series elastic actuators. IEEE/ASME Transactions on Mechatronics. 2020 Nov 6;26(1):586-91.Direct Empowerment for Joints Weakened By Limb-Loss
The medical community classifies persons with amputations into “K-levels” according to their athleticism and confidence in maneuvering with their prostheses. While the most capable K4 ambulators dominate the subject pools for engineering experiments in powered prosthesis, the K2 “community ambulators” may not be able to fully benefit from the powered-knee and powered-ankle configurations of modern devices such as the Open Source Leg. The reason is that many K2 ambulators have a shorter residual limb, and therefore have lost significant muscle mass in the hip-actuating muscles that attach to the thigh. But these hip muscles are critical to actuating the biological hip and effecting the motion of passive prostheses and especially the inherently heavier powered prostheses. This presents an obstacle to effective robotic intervention and impedes K2 ambulators from activities of daily living, negatively impacting their quality of life. Recently, Ref. [IshmaelTranLenzi2019ICORR] proposed to address this problem mechanically, by fitting patients with an under-actuated orthoprosthesis—a powered hip orthosis connected to a passive prosthesis—and applying assistive torques to the weakened hip. However, the control strategy remains task-dependent, since it assumes that the person is moving periodically, and is thus unable to assist in the full range of non-periodic behaviors included within activities of daily life. These include sit-to-stand maneuvers, transitions between tasks, and recovering from unexpected disturbances in order to avoid falling. The overall goal of my research thrust in this area is to extend this mechanical assistance strategy to a general class of under-actuated orthoprosthesis feedback systems for task-invariant assistance of joints weakened by amputation surgery, with initial emphasis on K2 transfemoral amputees. I hypothesize that by implanting a six axis force/torque sensor in the passive prosthetic, just below the socket, the biological hip torque can be both estimated to a high degree of accuracy and fed back via the powered orthosis to increase the wearer’s perception of physical strength at the hip, thus facilitating their use of the prosthesis. More specifically, I aim to (1) design and assemble a specialized under-actuated orthoprosthesis that features an instrumented passive prosthesis, (2) understand the dynamics of the residual limb of K2 ambulators for the purpose of guaranteeing stability under feedback with a hip exoskeleton, and (3) determine the extent to which hip strength amplification changes user perception of the mass and inertia of the passive prosthesis while walking. As this research addresses a key technological barrier to biomedical adoption of an assistive exoskeleton device, it would align with the mission of the NIH NIBIB. It also aligns with the NIH NCMRR’s Devices and Technology Department and research priority in prevention and treatment of secondary conditions. Follow up investigations could introduce hardware for amplifying other affected joints and body-powered hand prostheses.
Estimation-based control of powered prostheses
An estimated 2.2 million Americans are suffering from limb loss today [ZieglerEA2008APMR]. With conventional passive prostheses, people with lower-limb loss often struggle to perform the various activities of daily life since these devices cannot recreate the full energetically active function of the missing limb. Powered prosthetic legs offer a partial solution, but state-of-the-art control for prosthetic legs almost universally switches between states of a finite state machine, a strategy that both requires elaborate user-specific tuning and is inherently task-dependent [TuckerEA2015JNER]. An emerging alternative control approach, which I co-developed with Dr. Gregg in the context of exoskeletons, uses dynamical system state-estimation to overcome this task-dependency and allow an estimator to continuously track shifting locomotion tasks in real-time. The method extends Dr. Gregg’s prior approach [RezazadehEA2019IA] by fusing all possible sensors rather than the thigh angle alone and by estimating continuous task variables as the stride is in progress rather than after the stride is over. This dynamical system includes a single “gait-phase” oscillator as well as a bank of gait parameters ranging from concrete measures like the ground inclination to abstract descriptions of what differentiates one stride from another. These descriptions could encode, for example, a “gait-compensation-fingerprint” that describes how an individual stride has been volitionally altered from the baseline pattern.
Such compensations could then be mapped to degrees of freedom in the powered prosthesis that the patient would otherwise be unable to control, such as swing height or propulsion during stance. More specifically, my aims are to (1) model a practical subset of the possible modes of human locomotion as the oscillation of a single, fundamentally low-dimensional dynamic system using a bio-mechanical dataset of able-bodied human locomotion, (2) apply nonlinear state-estimation to estimate the low-dimensional dynamic system’s state in real-time based on all the sensor information available to the prosthesis,
and (3) determine the extent to which participants can control the device when this gait-state estimate is mapped to the “corresponding” robot behavior. As this research addresses a key technological barrier to biomedical adoption of powered prosthetic devices, it would align with the missions of both the NIH NIBIB and the NSF CBET program on Disability and Rehabilitation Engineering.