Walking faster and further with AI super-boots
AI is enabling the next generation of technologies which will help people to walk faster and with less effort — helping the elderly, those with muscular weaknesses and the ubiquitous student running late to class alike.
One salient example I wanted to share relates to exoskeletons that turbocharge walking. An exoskeleton is a device which attaches to the body externally and enables mobility in some way — the ones focused on walking tend to look something like a boot.
Years of research have gone into developing exoskeletons which create ease in walking — they do this by adding a push with every step, increasing the rotational force (called torque) with which one walks. One of the bottlenecks in scaling these solutions is that walking is an incredibly personal act: everyone walks with an individual gait. As we walk, we pick up and slow down our pace repeatedly in our own personal way too, so if a device can’t be personalized to each person that wears it, it’s not very useful. In most cases today, the personalizing of these devices is done in a lab across multiple lengthy sessions.
A Stanford lab is tackling this problem by deploying an AI model onto an exoskeleton boot. In a nutshell, the model starts with some default settings, and as a person begins walking in the shoe, the model learns what works best torque-wise for the subject at different speeds and paces the person is likely to walk at on an average day. They are able to do this because of data collected on sensors in the boot and the ability to deploy this AI model in electronics on the device.
Imagine a future where everyone can walk at twice their fastest natural pace when desired, where as we age we have more support and much longer before needing additional aids like walking sticks, and rehabilitative machines adjust to us in order to help us build muscular strength or recover from injuries. Additionally, the untethered training technique is likely to translate to broad human-robot interaction contexts, providing a technique for personalizing means of interaction between a robot and human. For now, this device is a research prototype and not ready for commercial use, but the ability to optimize a personalized aid untethered is a huge success story on the journey to scaling the personalized mobility devices of the future.
Curious about how the AI works? Read on.
The AI model which the researchers deployed is one of the first you’re likely to learn about in any machine learning course — a logistic regression algorithm.
Logistic regression is a type of AI model which helps to choose the best category from multiple options. Given k (multiple) options, the model uses the data it has been given to choose the best of the k options according to a target goal (e.g., minimizing or maximizing an objective).
In the context of this research study, the objective is to minimize the estimated metabolic energy used to walk. This goal is measured through an estimate that relates data on walking like torques and ankle velocity to metabolic rates. Before deploying the model, the researchers trained this estimate on lab data from studies of tethered participants.
After the model is set up, it is deployed onto the exoskeleton device. When a new person puts on the boot and begins walking, the model starts with k default options which are k different combinations of factors related to walking like ‘ankle angle’ and ‘ankle velocity’. On each step the model uses sensors to see how different the person’s natural angle and pace of walking are to its default factors in the k different options. The model identifies the option which, taking into account the person’s natural gait, likely requires the lowest metabolic cost, through performing a computation which ranks how much each option is likely to reduce metabolic rate.
The model then performs an optimizing step, which updates the existing options a bit closer towards this best option, which creates k more options to be tested that reduce metabolic rate more but still leave room for further optimization. This process repeats itself until the objective is no longer getting significantly closer to the desired goal. This process of updating options to minimize a goal is often referred to in machine learning as gradient descent.
A high level diagram explaining how the optimization works is below, heavily referencing this Nature paper.
One important note is that as humans we walk at a range of paces — often speeding up and slowing down, and choosing different paces based on how much of a rush we are in. To handle this, the exoskeleton has to be able to adjust to handle multiple different speeds of walking. Therefore, the optimizing approach is actually done multiple times for different ‘speed bins’ (i.e. different speeds that the person walks at, in order to provide a speed-adaptive exoskeleton that is helpful regardless of what pace you are walking at in a given moment. )
What applications excite you about this technology? What do you want to hear more about related to this or similar AI applications? Excited to continue to unpack other examples of AI being used to power mobility and assisting people.
All opinions expressed are my own, and not of my employer or external affiliations.
