Off-device will dominate the factory and home. Excited to share my thoughts with @SemiAnalysis_ on compute in robotics as deployments scale. Culmination of supply chain, our work at @EfferenceAI, and conversations with the top robotics teams. Link: open.substack.com/pub/semian…

Without a doubt: A world running foundation models on a billion robots won't be doing it locally. None of the embedded compute will look like it does today. All of the embedded compute will be wireless-optimized. Here are the reasons why: (1/5)

I left my PhD program at Princeton, went missing for the better part of the past year, and built @EfferenceAI — perception and compute designed specifically for data-driven robotics. We're now aggressively scaling to more than 20,000 manufactured units/month and beyond and doing so with an amazing team. I knew nothing about supply chain and manufacturing when I started, and bear many battle scars that remind me of my naivety: > Bought 65,000 units of the wrong DRAM because I misread a data sheet (aka a lot of $$) > Skipped a critical testing phase and blew up a huge batch of PCBs (aka a lot of $$, again) > Locked in a large volume for a component with falsely reported specs on their public release (aka a lot of $$, AGAIN) But aggressively scaling hardware has its advantages… It’s hard to buy 20,000 of anything in robotics right now, and we’re one of the only American companies that can actually get it done. The mandate we are building for: The past 20 years of data-driven robotics required smaller volumes —> selling a few units to tens of thousands of customers… The next 20 years of data-driven robotics requires the opposite —> selling tens (or hundreds) of thousands of units to a few customers. If you’re in building in data-driven robotics, or in the market for some DRAM (plz buy), would love to hear from you.

Reliable self-driving for hundreds of $$ (cameras) instead of thousands (LiDAR). We’ve been testing depth maps and object detection at 50m and beyond. Videos show @EfferenceAI with D435i mounted on the front, side, and rear while driving around Russian Hill / Pacific Heights

These are point clouds from: functional-manipulation-benc… D405 cameras are EVERYWHERE in robotics (commonly wrist mounted like what I have on my YAM arms) but depth is rarely used because it’s not very good Our models make training 3D diffusion policies possible W/O the use of LiDAR

@EfferenceAI is building robotic cameras because nothing on the market does the things that make 3D vision efficient and effective in humans. Data representations matter (arxiv.org/abs/2502.20237) and humans have a lot of important design features (karger.com/bbe/article-abstr…). A rich and scalable representation for visuomotor policy learning will come from hardware that is inexpensive and computes high-quality geometric information directly on the PCB, which is exactly what we are building.