bexo V2.5

Active Back Exoskeletons for labor intensive workers.

A research project under EPIC Lab at Georgia Tech.

Lifting and loading tasks in the workplace are ​common sources of musculoskeletal injuries. Back ​Exoskeleton (Bexo) research team evaluates the ​performance of device during various manual tasks ​and how it might be integrated into workplaces to ​offset injuries.

Since I got to Georgia Tech in August of 2021, I ​took over the research project and have been ​leading the development of the mechatronics side ​of the device and also highly involve in the ​controller development process, especially low and ​mid-level controllers.

A system for actuating cables was chosen to optimize the ​torque exerted on the user's intervertebral and hip joints. This ​design aims to minimize user discomfort and preserve the full ​range of motion, despite the presence of the device.

This picture is a great diagram of how Bexo is able to ​provide assistance to the user during lifting motion. For each ​independent cable-actuation system, the cable is connected ​between the actuator located above the lumbar region of ​the user’s back and the opposite shoulder, routed through a ​pulley on the back of the user’s thigh.

This “X”-shaped pattern allows for asymmetric torque ​assistance across a range of lifting motions.

MEchanical Design ​(Bexo V2.5)

To understand the significant of V2.5 design, I will have ​to explain what V2 was like. V2 was quickly designed by ​the previous project lead before he graduated, just right ​before I got to Georgia Tech. There were a lot of design ​flaws that had not been worked out at that point. For ​example, the design was very bulky for a wearable robotic ​device, the assembly process was next to impossible, and ​most parts are very hard to manufacture and had a lot of ​weak points.

After doing the analysis on the design and considering ​the design specification, I decided that a design overhaul ​had to be done with the limitation of keeping all the same ​components.

The design still works as intended, but the device was ​very hard to maintenance and uncomfortable to wear for ​a long period of time.

Having all the design requirements in mind, I went to work ​on designing the V2.5.

V2.5, with 2.66kg, has a 34% weight reduction compared to ​V2 and better weight distribution. Every subject who has worn ​both version all said that it was a lot more comfortable.

From the engineering standpoint, this design was simplified ​down a lot while keeping the same device performance ​(proven by our later human subject experiments). Having a ​simple design means the cost of the device is lower as well. ​The device is also easier to assemble and maintained.

Simplifiled Pulley Design

An Example of the new design

After more usage and experiment with the device, we found design flaws that I have to keep re-iterating ​and coming up with innovative mechanisms to solve each issue. Obviously, we cannot go into every single ​fixes, so I wanted to highlight a few major ones.

Cable guide

Cables slip to the side of the body quite often when Bexo provides assistance during squatting ​motion. Therefore, cable guides have to be implemented to fix the issue

Spring loaded housing pulley

After certain hours using t device, the cable rubs ​against the pulley housing enough that it starts to cut ​through the printed plastic parts and starts to get snag ​on it. In addition, when a higher speed is commanded ​the cable may just completely tare through the housing. ​I could not find a picture for it, so I created this ​illustration instead.

Of course, changing the material to metal could fix the issue, but ​that would add a significant amount of weight, and the cable ​might be the part that snaps instead. The solution that I came up ​with is quite fun! To absorb the energy that the cable is tugging on ​the housing, I decided to add another translational Degree of ​Freedom (DoF) to the housing. It is spring loaded to help make the ​output hole not move around too much, so it has to reach a ​certain thread hold in order for the output hole to move. The ​strength of the springs can be adjusted using the screws on the ​side.

To make sure that the springs added doesn’t affect ​the output performance, I set up a benchtop testing rig. ​(Design Validation!) The setup comprises of AK80-6, ​the same actuator as the one on Bexo, 6-axis force ​sensor, and the new pully design. I basically just ​command a certain torque, ranging from low to high, ​and compare to the output torque (calculated by using ​T = Fr). The result showed negligible loss comparing the ​commanded and the output.

Rail System

We were toying around with two sets of electronics for a while during the controller transition period which I will ​talk about in a bit. Basically, we needed a way to switch around between two setups quickly. So, I created a ​modular rail system for Bexo!

This simple rail design was created and attached to the ​main body of Bexo. The carrier simply slides in and get locked ​in with a M3 screw at the bottom

HArdware Design ​(Bexo V2.5)

Not a lot to explain here, haha. The hardware set up here is fairly simple. Bexo has three main parts, the ​brain or controller, the muscle or actuators, and the feedback or loadcells. With the modular system, I made ​a sharable loadcell amplifier that can be connected to both controller modules shown in Rail System section.

AK80-6 or the muscle for Bexo uses CAN protocol which will be talking to myRIO in the legacy setup and ​Teensy in the new setup. We used SPI to CAN for myRIO since it doesn’t have a native CAN port, and we had ​to use a transceiver since Teensy doesn’t have one on board.

Everything is made with through-hole components and hand-soldered by me. I think the setup is fairly ​clean for how many components we have!

Controller Design ​(Bexo V2.5)

My main contribution to this part lies mainly in low and mid-level controller on Teensy 4.1 since I am fluent ​in C++. I have been dipping more and more into high-level controller (Machine Learning in this case).

Low-level code is super basic. It handles the ​communication between devices and the safety features! ​Since we are working with a device that is worn by human, ​we have to be absolutely sure that it will not harm the user ​if something goes wrong.

To identify what safety features I can implement, I use ​Failure Mode - Root Cause - Risk Mitigation Control method ​I picked up from my internship! After brainstorming, I came ​up with the chart you see here.

We had problem when we start communicating with ​other devices - it brings down the frequency to around ​50Hz. Teensy 4.1 has only one core so I couldn’t implement ​multithreading. To fix the speed issue, I had to implement ​something similar to hyperthreading by Intel on Teensy to ​bring the frequency up to 250Hz. The reason why the ​controller has to be running at a really high speed is ​because Bexo has a cable driven design, making precise ​control real-time harder if PID controller is unresponsive.

That brings me to mid-level controller which ​handle the PID controller. Bexo uses velocity PID ​controller to control the tension in each cable. ​The PID controller takes loadcell data in as an ​input that low-level controller collects and ​applies Chebyshev’s and Low Pass filters.

The bigger piece of the puzzle in this controller system is ​high-level controller. There are two controllers that we are ​developing right now. Impedance State Machine (ISM) and ​Machine Learning Controller (MLC).

For ISM we use IMU’s Euler angles and accelerations to ​determine the states - resting, going down, lifting. For each ​state, Bexo commands different cable tensions to make ​sure that the cable is sufficiently drawn tight in resting state ​ready for assistance, provided dampening on the way ​down, and given desired assistive force during lifting. This is ​simple and reliable.

However, we want to explore the idea of predicting ​biological torque at L5/S1 spine joint using MLC. We use that ​data we collected in summer 2022 to train a TCN model. As ​of this moment we have not tried it out on human since we ​have been trying to fix the messy data. Now, I am not going ​to go into details, but the collected data was really messy ​and completely wrong in some part. We had to clean the ​data through a bunch of processes to get the results you ​see here!

The graph you see is the 1 life cycle that our MLC ​predicted compare to a study. Looks real nice if you ask ​me!

Human Experiment ​(Bexo V2.5)

Working with exoskeletons involves testing the ​effectiveness of the device. So, to scientifically validate ​the exoskeletin, we use sensors like Electromyography ​(EMG), metabolic, and motion captureing.