ATOS is the Assistive Technology Operating System. It turns the powered wheelchair from a closed, single-vendor appliance into an open, modular platform you actually own. Apache 2.0 and MIT. Forkable. Auditable. Un-discontinuable.
A wheelchair should not have a date of death.
A five-year, $41M ARPA-H program. Eight institutions. One open platform.
Two companies, R-net and LYNX, write the software inside nearly every power wheelchair sold in the Western world. You cannot read it. You cannot change it. You cannot add to it. The thing that is functionally your lower half runs firmware you are not allowed to open.
Every manufacturer has its own proprietary messages and handshakes. Nothing talks to anything. Want to add a sensor, an arm, an AI assistant? The controller will not let you. Not because the physics are hard. Because the garden is closed.
The Pentagon has a name for this disease: the cost of a tightly coupled system where nothing can be replaced without re-validating the whole. A cockpit screen that costs $10M to swap. A wheelchair has the same disease, on a larger human scale.
When the company that built your chair sheds product lines, you learn a quiet kind of panic. Invacare filed for Chapter 11 in 2023. Your iPhone going under means a new phone. Your wheelchair going under is welded to your body.
A closed system cannot evolve. It can only be replaced. That is what we are building. The replacement.
It boots. It reads. It writes. It updates. It outlives the company that shipped it. Unremarkable in a phone. Revolutionary in a chair. ATOS gives the chair the verbs the phone got twenty years ago.
Perception, planning, manipulation, safety, UI: independent enough that any one can be replaced without rebuilding the whole. A better algorithm out of CMU next year? Pull it in.
Ship the core once, then load and swap motor-control behaviors and apps in the field. No dealer visit. No service van. No permission from a man with a cable.
Dual-licensed Apache 2.0 and MIT, the pair Rust ships under. The firmware lives on a server somewhere, mirrored, forkable by anyone with a laptop and a need.
Standard joystick, sip-and-puff, eye-tracker, or a BCI implant. Any assistive input is a plug-and-play module on a standardized communication layer.
A university ships a tremor-cancellation plugin. A company sells smart-home integration. Discovered, certified, and installed without bricking the drive system.
Protocols published license-free. Anyone with the skills can diagnose and fix the chair. No more waiting weeks for an expensive part from an authorized vendor.
ATOS is mixed-criticality. Work is split across domains so a fault, or a slow AI workload, can never compromise the drive and braking that keep a person safe.
Untrusted, portable third-party plugins, custom UIs and novel algorithms. Memory-isolated from the drive system, language-agnostic, binary-compatible across ARM, x86 and RISC-V.
Navigation, perception, mapping and manipulation on a hypervisor + Linux (e.g. Jetson). Where the intelligence lives, isolated from the safety domain below it.
Joystick, e-stop, motor control, braking and hard limits on a microcontroller. A safe-state default, watchdogs, fault containment, and limits the core enforces that no module can ever override.
Two loadable-module mechanisms, two layers, complementary not competing: native elf_loader modules
for near-bare-metal speed in the deterministic loop, WebAssembly for safe, portable apps above.
Shared autonomy is a spectrum, not a switch. The technology's job is to read where you are on the scale between I want to do this myself and I want the machine to do it for me, and meet you there.
You steer with the lip joystick. The system smooths tremor and warns before a curb, but every decision is yours.
Autonomy that removes agency is not assistance. It is paternalism with better sensors. The user has to own the decision, or the whole point is lost.
RAMMS renders the chair, every sensor, every motor, every joint of the arm inside Unreal Engine 5 at frame-accurate physics. Test the whole stack in software before any of it touches a real person.
And because it is a photoreal world with a physics-accurate chair, it is also a game: the research version trains the AI, the clinical version trains the user, the consumer version teaches an able-bodied player what a world built for someone else feels like.
The bottleneck in embodied AI is data. There is no ImageNet for a hand reaching for a cup. But a power wheelchair already moves through every environment a human body must: kitchens, bathrooms, sidewalks, elevators, the back of a Lyft. It is the ideal sensor platform for physical AI.
Each chair trains locally, on its own environment, with its own user. Your routes, your routines, your preferences: nobody else's business. Only the learned improvements get pooled across the fleet. Collective intelligence goes up. Individual privacy stays intact.
A user in Omaha ends up teaching a user in Boston. Not directly. Just by living. The first wheelchair user to solve a particular refrigerator leaves a path for everyone who comes after. You would not be alone inside your chair. You would stand on the shoulders of everyone who drove one before you.
The disabled body is the QA department for the species. The data has been there the whole time. Nobody bothered to collect it.
Not a triumph. Not a story about overcoming. A morning where the technology disappears so completely into the background of a life that you forget you are operating it.
Every step an able-bodied person performs without thinking is, for me, a request, a coordination, a transaction. Those numbers are not a complaint. They are a specification.
A robot you notice is a prototype. A robot you don't is the future. Time is what dignity actually measures.
ATOS is being built under RAMMP, the Robotic Assisted Mobility and Manipulation Platform: a five-year ARPA-H program that funds milestones and products, not just papers. Governed in public by an Open Source ATOS Foundation so the software a person depends on to move cannot be discontinued by any single company.
Principal Investigator, RAMMP · Human Engineering Research Laboratories, University of Pittsburgh
Thirty years studying wheelchairs and the people who use them. A wheelchair user himself, and a 1988 Paralympic bronze medalist. He treats lived experience as data, not decoration. He has been waiting for the industry he helped build to finally be ready to do this.
The company can die. The engineers can move on. The venture capital can dry up. The repository is still there, mirrored to a hundred servers, forkable by anyone with a laptop and a need. That is what open source actually means in a life-critical context.