Essential tremor (ET) is a movement disorder characterized by kinetic and postural tremor that affects an estimated 7 million people in the U.S. alone. Deep brain stimulation (DBS) of the ventral intermediate (VIM) nucleus of the thalamus is an established medical therapy for the treatment of ET and has been shown to decrease tremor symptoms significantly for patients in which traditional pharmaceutical therapies have proven insufficient. There are limitations to continuous DBS treatment however, with battery replacement surgeries necessary every few years as well as the presence of side effects such as dysarthria, paresthesia, and gait ataxia. Adaptive DBS (aDBS), with its use of internal or external markers as feedback to modulate stimulation parameters, presents a promising avenue through which to mitigate these concerns. However, current aDBS methods typically employ a binary, “on-off,” control system, which may introduce new complications to consider. One such complication is rebound effect, a transient increase in tremor severity immediately following the deactivation of DBS before levelling out to a steady state. However, clinical observations of rebound in ET have been mixed in the literature, which results in an ambiguity about the potential for effect on an aDBS system. To clarify this mixed literature and quantify the potential impact that the rebound effect may have on an aDBS system, we collected inertial measurements from the tremoring arm during the spiral and line drawing tasks of the Fahn-Talosa-Marin tremor rating assessment task. This task was used to characterize rebound over a ~40 minute time window, over which we observed the rebound effect in all three of our patients. Our results indicate that rebound effect should be taken into account when designing aDBS systems, pointing towards the development of a non-binary aDBS system.
 

The world’s largest high-energy particle accelerator, the Large Hadron Collider, relies on sub-millisecond processing performed on massive amounts of data coming out of its particle detection system, which is due for a major upgrade in 2024. The old particle detection chips will be upgraded to RD53B chips with faster data transmission, allowing for more complicated data processing. The readout systems that interact with the particle detection chips, YARR and FELIX, need to be tested with the RD53B chips and debugged before the system is put in place. The goal of our research is to create an emulator of the RD53B chip that can produce dynamically generated pseudo-realistic data at the same rates that would be seen in the Large Hadron Collider without the need for heavy radiation. This will allow for readout software to be fully functional and debugged before the actual RD53B chips are fabricated and placed into the Large Hadron Collider. We are using Field-Programmable Gate Arrays (FPGAs) to mimic the hardware inside the real RD53B chips. In place of RD53B’s analog sensors, we have substituted digital logic that generates pseudo-realistic data because FPGAs cannot emulate analog hardware. The alpha version of the RD53B emulator with basic communication and pre-programmed data was completed in February. Recently, the beta version of the emulator with dynamically generated data was completed, and we have been testing communication between the emulator and FELIX. With the beta version of the RD53B emulator tested and verified by us, the developers of YARR and FELIX will use our hardware to help verify that their systems will provide accurate readouts from the real RD53B chips. The next steps for the RD53B emulator include a hardware data decompression accelerator, as well as any additional features requested by the YARR and FELIX teams.

The mobility of autonomous walking robots is an essential characteristic in their operation. Due to currently imposed constraints in battery technology, the optimization of robotic locomotion for energy efficiency is paramount. Previously, elastic payload suspension has been employed to reduce the cost of transportation in a hexapedal robot. Prior results suggest that the optimal load suspension characteristics are a function of robot morphology and locomotion strategy. A payload suspension system that can be easily adjusted would allow for the accommodation of a variety of these morphologies and locomotion strategies. We have designed a tunable suspension system that will allow for experimentally determining optimal suspension characteristics in a cost-effective manner. The design enables continuous adjustment of the suspension stiffness and damping, so optimal parameters can be determined through hardware experimentation. This hardware experimentation allows for the creation of a numerical model for an oscillating payload’s behavior, which can be compared to simulations. We have completed calculations and design of this hardware, and anticipate seeing that the data collected from its usage will verify the utilization of a haptic testing system in robotics development and allow us to determine methods for calculating the best parameters for elastic payload suspension. This verification of simulated data will allow for further research in developing more efficient methods of payload attachment to legged robots, examination of locomotion when carrying payloads, and design of payload management methods.

The relationship between humans and robots is full of feedback loops: how our brain processes what we see, feel, and act affects how the intelligent machine reacts and vice versa. Understanding this feedback loop will enable us to design better day-to-day automated systems. For instance, operations performed with surgical robots are guided with the surgeon’s movement. My research in the UW BioRobotics Laboratory investigates how the human brain adapts to technology. I programmed a haptic paddle to apply timed forces on human subjects to investigate how uncertainty, in the form of a disturbance, affects response over time. Understanding how human behavior changes is crucial for improving automated systems that can be simulated by the haptic paddle such as cable-driven surgical robots. We apply the haptic paddle to model forces a surgeon may experience during surgery, such as tough tissues or other disturbances, while operating a robotic surgical device. As we measure the participant’s learning curve over longer periods and more trials, the participant builds a better understanding of how to react relative to their force exerted on to the paddle. By increasing our understanding of how the human brain works, we can begin to improve precision of surgical operation.

After neurologic injuries like stroke, users must relearn how to interact with devices to achieve different tasks. However, hand dominance before and after neurologic injury could affect motor learning and recovery. Understanding a) whether hand dominance plays an effect on motor learning and adaptation of predictive (i.e. feedforward) and reactive (i.e. feedback) controllers and b) how motor learning is impacted by which side was affected by stroke is crucial for personalizing rehabilitation techniques. We are currently investigating the effect of hand dominance on motor learning during continuous tasks for users without motor impairments. 10 participants were asked to play a simple trajectory-tracking game, first with their non-dominant hand and then with their dominant hand. We found that participants’ feedback controllers improved with practice for their non-dominant hand (p=0.005), but their feedforward controllers were unchanged (p=0.88). Furthermore, the participants transferred their feedback controller from their non-dominant hand to their dominant hand (p=0.33). This suggests that the non-dominant hand may be more specialized for error correction and impedance control, and that rehabilitation for users who had a right-hemisphere stroke should focus on improving reactive control.

Filtering the data produced by the Large Hadron Collider (LHC) is computationally challenging due to the sheer quantity of the data, on the scale of hundreds of terabytes per second. In the coming years, data production for the LHC is projected to increase by a factor of 15 with the high luminosity upgrade. Machine learning algorithms could provide pattern recognition capable of filtering data produced by the LHC. Specialized hardware could increase the throughput to match the data rates required by the LHC. We analyzed several cloud-based specialized hardware solutions including Amazon Web Service FPGAs, Microsoft’s Brainwave Service, Google’s TPU, and NVIDIA GPUs to compare the performance of each of them for particle physics application. The networks accelerated were trained on a variety of data including: Top vs QCD quark classification, Hadron calorimeter data, and electron energy regression. These experiments demonstrate the feasibility of machine learning algorithms in high throughput required situations such as high energy particle physics.