Work Experience

Papers using the Samsung GEM exoskeleton suggested a lighter, lower-torque design could deliver more metabolic benefit than a heavier, higher-torque one. Hypershell's commercial device weighs just 2.5kg without batteries — light enough that testing this hypothesis, before committing the lab to a full from-scratch build, seemed worth doing first.

I designed an adaptor to mount our own motor and a custom PCB (mounted on top of the Jetson) onto the Hypershell structure, while reverse-engineering the device itself from the physical object — no manufacturer CAD or drawings existed, and Hypershell's thigh interface has an unusually complex curvature that made the reconstruction hard.

Chasing a flickering reading

During testing, the motor encoder reading was intermittent — good sometimes, bad other times.

My first hypothesis was the mounting material: the lab was using PLA at the time, which deforms under motor heat, so I respecified it to PAHT-CF. The reading was still bad. The real cause turned out to be that the driver board couldn't be bolted down without matching the commercial unit's own hole pattern, which no drawings existed for either — solved with a custom shadow-casting rig, projecting the part at a larger scale to measure the hole coordinates on a lathe. About two weeks of work in total.

With a reliable reading, I ran a metabolic test on the converted device: a 5% reduction versus no exoskeleton, single subject — not conclusive, but the first measurable metabolic benefit this lab had produced. The past student's device hadn't shown one. That result was enough for our PI to commit to a full reverse-engineering of the device.

Full reverse-engineering began at speed — I wanted a testable result fast, so I started with whatever was quick and cheap, evaluated it, and moved on when it failed. My CAD ended up visually near-identical to the commercial original, built entirely from measurement since no manufacturer drawings existed. Our 3D printer was too small for the full-scale parts, so I redesigned around that constraint while keeping the structure intact.

See full-size render ↗

When motor rocking showed up under load — the commercial device itself already rocked at 9 N·m — I had three suspects and worked all three at once: the pelvis interface, insufficient thigh contact, and the hinge.

The pelvis interface

Hypershell's own pelvis part sits on top of the glutes rather than the lower back, which I suspected made it inherently less secure. I tried several back-plate drafts — deliberately cheap, fit-check only, never load-tested for torque — but none fit perfectly. This thread stayed open through the end of Mark I.

The thigh interface

I considered adding high-thigh straps for more contact area, reasoning more contact would reduce rocking. Weighed against the comfort and gait trade-offs, I set the idea aside for the time being — a thread I'd pick back up in Mark II.

Hinge failures

The hinge looked like the fastest fix, so it got most of my attention. The door hinge already in place had gaps both between the pin and leaf and leaf-to-leaf. A quick plastic hinge sourced off Amazon didn't endure even 12 N·m.

Meanwhile, a McMaster metal hinge held to 12 N·m but rocked worse than the plastic one, from the same leaf/pin gap problem.

I modified it myself — dowel pins, redrilled holes, shims — which fixed the wobble and pushed it to 15 N·m, where it broke. Three off-the-shelf hinges, three separate workarounds, still failing: that's what convinced me to design a custom hinge from scratch instead.

By the end of Mark I it was already April — seven months into the project — and I still didn't have a system working at high torque. I concluded the block-like pelvis design simply didn't fit well across different people, while Georgia Tech's more ergonomic approach looked promising: more pelvis contact points, better support. Test prints confirmed it fit well, and I committed to a Georgia-Tech-inspired pelvis interface from that point on.

I also made a deliberate scope cut: stop trying to fit everyone, and focus entirely on getting one working system — on myself. That meant abandoning some Mark I generalization work, like an extendable pelvis meant to fit a range of hip widths.

Building the ergonomic pelvis

With no body-scanning equipment available, I measured myself the old-fashioned way — circumference and midpoint-to-endpoint distances across the upper thigh, hip, and lower back — and used SolidWorks surfacing tools to reconstruct a model of my own pelvis, glutes, and thigh region from those measurements, then built the pelvis interface directly off that surface, crediting Georgia Tech's device design for the underlying approach. It fit me perfectly. It also surfaced a new problem: the resulting body-facing surface isn't flat, but the motor needs to sit flush on the sagittal plane or it delivers torque at an angle. Not severe enough to block progress yet — I didn't solve it until Mark III.

Eliminating the hinge

Motor rocking still hadn't gone away, even with a good pelvis fit. Having already lost faith in off-the-shelf hinges, I built the ABAD joint directly into the pelvis interface with no separate hinge at all, which eliminated hinge wobble as a variable outright. (Every ABAD hinge through Mark I had sat below the motor — this was the first time the joint moved close to it.)

Why distance from the motor matters

Rocking still persisted, which sent me back to the thigh interface as the last suspect. I think of it as a link–motor–link problem: secure both far ends of a link and the motor in between has room to rock; secure the link close to the motor instead, and it should rock less. Two things likely explain why. Any small angular play at a loose joint — a sloppy pin, an under-constrained contact — produces a linear displacement proportional to how far that joint is from the point you're watching, the same reason a wrench wobbles more at the far end of the handle than right at a loose bolt head. And if the structure itself flexes rather than just having play, cantilever-beam deflection under load scales with the cube of the distance from its support, so a compliant joint far from the motor bends disproportionately more than one close to it. Either way, constraining the far ends of the assembly — a good pelvis fit alone — wasn't enough; what mattered was constraining something close to the motor.

I came back to the high-contact thigh idea from Mark I, working through the earlier comfort and gait trade-offs by testing it directly. Using the same measure-and-surface method as the pelvis, I built a tightly-fitted thigh cuff leaving almost no gap against my leg.

Separately, my PI wanted the motor to be easy to attach and detach so it could be swapped across interfaces. I designed a removable hip-joint–pelvis connector for this — a shoulder bolt through the top of the motor that unscrews to remove the motor and hip-joint interface as a unit. That led directly to the idea that finally solved the hinge problem: since it already needed to be a shoulder bolt, swapping in one with a bushing turns the same fastener into a purpose-built hinge, solving the modularity request and the hinge problem with one part.

Tested in one attempt: no motor rocking at all, up to 17 N·m.

Not a controlled trial — Mark I is marching in place at 12 N·m, the same setup that broke the plastic hinge; Mark II is walking at 17 N·m, higher torque under real dynamic gait. If anything, that makes Mark II's zero rocking the harder result, not an easier one.

Solving the rocking issue brought the weight back up from 3kg to 4.1kg (with battery) — a byproduct of building tightly-fitted, integrated parts (the hingeless pelvis, the custom shoulder-bolt hinge, the close-contact thigh cuff) rather than a purely weight-optimized structure. I'm treating that honestly as a trade-off, not a regression. This is also a device custom-fit to one body — my own — which is its own limitation, and the subject of Mark III.

See full-size render ↗

By May, I wanted to spend more time on the control side of the project and stepped back from hardware, handing it to a PhD student in the lab. I only stepped back in when an undergrad's design attempt didn't work out.

The changes I made directly: reworked the pelvis interface to be movable, so it could generically fit a range of pelvis shapes across the lab — which immediately resurfaced the motor-mounting problem deferred from Mark II, since a varying pelvis shape made the non-flat mounting surface unavoidable. I fixed it by offsetting the motor for clearance. I also redesigned the strapping, from one stomach belt and one pelvis belt into two cross belts — noticeably lighter than before.

Straight bar, and cleaning up

An undergrad developed a straight-bar alternative configuration — a good concept that needed polishing, but he left partway through the summer before finishing it. I ended up cleaning up the CAD for the whole project.

Removing the hinge for good

The PhD student in charge took my advice to remove the ABAD hinge entirely — it does very little during level-ground walking anyway — replacing it with a carbon-fiber sheet that flexes slightly in the ABAD direction, closing out two generations of hinge failures for good. The PhD also tried a passive thigh-damping system; it didn't actually damp, but it did apply the strap-close-to-the-motor principle from Mark II.

See full-size render ↗

The result: about 3.2kg, tested on three subjects for metabolic benefit at roughly 10 N·m — 8 N·m for one subject, well under what Mark II proved the hardware could handle. That's the number that mattered: it's the direct test of the hypothesis that opened this whole project.