I’ve been obsessing over a concept for my startup for a while. Any true attempt would instantly be bottlenecked by R&D costs, but with infinite cash, I’d build something like the "Neural-Link Workforce."

Think: A cyber-physical bridge between untapped human potential and industrial robotics.

We have millions of people with physical disabilities who are mentally sharp but locked out of the physical workforce. Then there’s mass layoffs and economic disruption thanks to that misnomer, AI. Has anyone here ever wanted to build a high-fidelity, non-intrusive BCI (Brain-Computer Interface) through a VR headset medium and a line of workforce robots that humans can plug into and earn from?

Just me then?

The main idea would be, the user wears a headset and "jacks in" to a remote field robot (humanoid or specialised chassis). The bot would be running something heavy like an NVIDIA Jetson AGX Orin to handle the "reptilian brain" tasks locally. This manages the inverse kinematics, dynamic balancing, and SLAM, so the latency doesn't kill the experience. The human pilot would provide the high-level intent, strategy, and problem-solving. It’s basically monetising telepathy.

If the budget is truly unlimited, then we’re not talking about some Circuit Mess build. You’d be looking at a custom HMD integrating a high-density active dry-electrode EEG array. MAt be daisy-chaining TI ADS1299s for 64+ channels of resolution. Pair with fNIRS sensors to correlate hemodynamic response with electrical intent. We’d likely need an onboard FPGA (something like a Xilinx Kintex) just to handle the real-time artefact removal (filtering out eye blinks and jaw clenching) before the signal even hits the stream. All that good stuff.

I’ve crunched the napkin math on this. With an unlimited budget to kickstart the hardware, you could squeeze a sustainable model out of this in the near future:

1. The Pilot (User): It’s a Hardware ISA (Income Share Agreement). We don’t sell the expensive headset upfront. We place it for a deposit (or free), and the pilot trains in our VR "Academy." Once they get certified and land a remote contract piloting a bot, a % of their paycheck goes back to paying off the hardware. It effectively removes the barrier to entry.

2. The Employer: They can buy the field bots outright, or they subscribe to RaaS (Robots as a Service). They get the hardware and the maintenance; we’d supply the training and talent network via the Academy.

You could validate this by activating millions of previously unemployed individuals with a disability, but it could extend to the rest of the economy aa the tech evolves. Something like this banks on a hybrid solution to hedge against disruption. Instead of AI and robotics displacing humans, this concept makes the robot the medium of employment but keeps even the most disadvantaged empowered throughout the race. It could theoretically stabilise the workforce throughout any adoption phase by keeping humans in the loop. Machines do the heavy lifting, we do the thinking + oversight, and the economy keeps moving.

Isn't it what's already happening with some "autonomous robots" ?
electrek.co/.../

Haha yes! The difference I see though is Tesla attempted a Wizard of Oz reboot to sell a fantasy, whereas I’m saying we should just put that operator on the payroll without any illusions or fake hype. Selling 'full autonomy' right now is like selling commercial tickets to Mars before we’ve figured out cost-effective launch reusability or shielding passengers from cosmic radiation. It's just hype to sell pre-orders on distant tech that could change on a dime tomorrow.

I’m saying we more honestly approach the problem of autonomy while empowering real people (who should matter more in the economy than some glitchy clanker).

The interface tech I’m advocating (BCI/VR) work now. We don't need to wait for a sentient biped. We could theoretically plug untapped labour into existing industrial units, exoskeletons or drones tomorrow. Then we just wait to see how Physical AI plays out before we hybridise it with HitL (that is, if investor capex into transformer architecture hasn’t formed a giant bubble around current gen AI). That way Human-in-the-Loop is the feature, not the bug.

This is definitely something that could grow legs lol. I had also considered what the industry calls Sim-to-Real Learning from Demonstration (LfD) to effectively plug more advanced pilots/engineers into a digital twinning system and run parallel optimisations in real-time. This is effectively RLHF on steroids. That’s far more scalable than only offering manual or hybrid piloting too, because now the teleoperator isn’t constrained to one given role or job, and the system ultimately maintains their economic leverage as the enabler that can better translate higher-level business logic to low-level robot logic.

With that said, embodied AI then becomes something of an assistive technology in the human feedback loop (and vice versa). To derisk, maybe a Diffusion Policy (like in generative art) can take the noisy/messy human inputs and denoise it into a smoother (mathematically perfect) outputs that capture the intent of the operator. With such a model, you could capture this intent such that a physically-disabled/non-technical pilot can still effectively code in C++ or Rust without writing a line of it, therefore optimising a workflow in real-time with all the necessary telemetry (and serving as a conduit between goals and action), simply by demonstrating intent via physical sim.

Obviously the digital twin here serves more like an IDE, and you would need to solve the big problems like world models / proprioception and even pain (as a cost function), likely leveraging JEPA, NeuroPINNs, state space modelling, etc.

it’s a fascinating thought that theoretically allows anyone to hit or say “Save" upon task completion, and the system just compiles their physical movements into a Neurosymbolic Policy that guides the robot code/kinematics remotely (like engineers on the ground floor already do). That way we wouldn’t be sacrificing robot autonomy and meanwhile keeping it all human-centric.

Thanks for the input!

DARPA's PAL and CALO projects were basically focused on AI that learns from the HitL until it was competent to do most of the work itself. I agree that's a good way to build up automatic robotic capability, though I'm not sure LLMs are going to be the necessary infrastructure for pure physically embodied AIs*. And of course there's the whole "train your replacement" issue of that approach, where more and more work themselves out of existing jobs. I'm not sure there will be enough (or any) "vibe coding" positions in 10 years to cover the labor surplus.

At least right now, jobs like plumber or electrician are pretty safe from automation, but that can change with enough observation and training given the ability to form higher level models from observational inputs.

[*]: OK, I'm personally pretty sure they won't be, but I've been wrong before.

Sounds like you are talking about more of an existing telechir type setup with the traditional user interface replaced by a more modern BCI/VR system.

Another take on this could be that you connect the BCI/VR system to a robot simulator to allow the user to program, edit, simulate, and optimise the code, which then gets downloaded into the existing production robot to then run repeatedly at speed with precision.

Sounds like you are talking about more of an existing telechir type setup with the traditional user interface replaced by a more modern BCI/VR system.

Another take on this could be that you connect the BCI/VR system to a robot simulator to allow the user to program, edit, simulate, and optimise the code, which then gets downloaded into the existing production robot to then run repeatedly at speed with precision.

This is definitely something that could grow legs lol. I had also considered what the industry calls Sim-to-Real Learning from Demonstration (LfD) to effectively plug more advanced pilots/engineers into a digital twinning system and run parallel optimisations in real-time. This is effectively RLHF on steroids. That’s far more scalable than only offering manual or hybrid piloting too, because now the teleoperator isn’t constrained to one given role or job, and the system ultimately maintains their economic leverage as the enabler that can better translate higher-level business logic to low-level robot logic.

With that said, embodied AI then becomes something of an assistive technology in the human feedback loop (and vice versa). To derisk, maybe a Diffusion Policy (like in generative art) can take the noisy/messy human inputs and denoise it into a smoother (mathematically perfect) outputs that capture the intent of the operator. With such a model, you could capture this intent such that a physically-disabled/non-technical pilot can still effectively code in C++ or Rust without writing a line of it, therefore optimising a workflow in real-time with all the necessary telemetry (and serving as a conduit between goals and action), simply by demonstrating intent via physical sim.

Obviously the digital twin here serves more like an IDE, and you would need to solve the big problems like world models / proprioception and even pain (as a cost function), likely leveraging JEPA, NeuroPINNs, state space modelling, etc.

it’s a fascinating thought that theoretically allows anyone to hit or say “Save" upon task completion, and the system just compiles their physical movements into a Neurosymbolic Policy that guides the robot code/kinematics remotely (like engineers on the ground floor already do). That way we wouldn’t be sacrificing robot autonomy and meanwhile keeping it all human-centric.

Thanks for the input!