
There's a lot of noise in the tactile space right now, and as someone who designed tactile sensors for prosthetic hands that actually had to work in the real world, I'm dubious of some of the more celebrated solutions.

physical AI is only robotics if it's truly physical

There's a lot of noise in the tactile space right now, and as someone who designed tactile sensors for prosthetic hands that actually had to work in the real world, I'm dubious of some of the more celebrated solutions.

looking for a cracked social media launch specialist for a physical product. Paid gig, knowing about robotics and physical AI is preferred but not necessary. Must be in SF, DM for details.

Everyone is talking about AI powered PCB design, maybe because it's easier to be impressed by shiny hardware. But I've been more impressed with AI powered firmware design. It is outrageously cool to receive a new, very dense custom PCB from your supplier, plug it into your computer, upload the PDF schematic to Claude, and come back 20 minutes later to the board fully brought up, all the LEDs working, switches working, Bluetooth working, profiles made, environment configured, etc. It ran through put tests, load tests, torture tests, etc. It hunted down all the well-regarded libraries for every digital sensor we used, and found bugs in two of them (both authored by tier one silicon companies). 30 minutes later, it had Bluetooth OTA updates working, and also informed me that we slightly undervolted our external flash chip, but that it was still working. 😅 And before anyone asks: the nice thing about firmware is that, when well-designed, can be just a few thousand lines. It doesn't (as easily) get out of control the way desktop/web applications do. Now, I'm still paranoid enough that I reviewed every single one of them, and...I do not believe I could have done better. Maybe with 2 weeks and a ton of trial and error. Ask any engineer about the time they lost eight hours over a typo in a data sheet or a race condition that only showed up when a component booted after being off for more than 30 seconds. The AI did make one mistake: it asked me for my opinion on how to pack our data into the wireless stream. 😅 Afterwards we got through that, it made several useful improvements. I jokingly told it to never ask for my opinion again. I panicked when I saw "memory updated". (Had to fix that). 😱 It is absolutely amazing to not have to add all the new age RTOS bloatware just to avoid hand-writing non-blocking timeouts all over the place, and to know that a system with 25, unblinking, all-seeing-eyes (all of them sharper than your 2) optimize the state machine to not wait around for some DRDY signal that ST insidiously made NOT show up every 20-50k reads. Human beings really suck at catching that stuff - you have to be a detective, and you end up writing one line of guard rail code for every line of actual logic. Of course we should wait before the stuff touches aircrafts or anything critical, but I believe eventually it will be far safer than a human. People denying the power of this technology are probably the same ones who still code in vim and hate compilers.

when did we start using all this weird jargon to obfuscate what we are doing in robotics??? You know who does that? Lawyers. Stock traders. Accountants. People who protect their trade by adding smoke and mirror complexity. Trades that don't eject you for faking it till you make it. Gold club trades. Most experienced engineers try to not do this, because engineering (as opposed to weekend warrior SF hackathon toying around) is hard enough without an added layer of bullshit. Real engineers are usually so cooked by the complexity of what they are doing that they (to an often frantic degree) insist upon simple and straightforward explanations. The whole idea is that you can all speak precisely about what things actually are - so you don't have to use a bizarre mash of made-up acronyms and re-branded old basic things. EEF? You write it once in a post (using AI haha) and you need to what, save time? Someone said to me "the EEF has a standard geometric reference point in our sim". Yes, it was called TCP (that's tool center point - grown ups speak plainly, enthusiasts gatekeep) for 1,000 years. We already had too many acronyms. It keeps going: -Robotics (the entire field) got turned into embodied AI/PI. -Teleoperation (tried to) turn to "human in the loop" (no idea who that brilliant marketer was) -UMI/VLA/VLM/WAM (these are okay - but frequently not rigorously defined when presented, it is a broad spectrum now, which means we need actual terms, not just branding) -Dangerously wrong slang (a guy was trying to tell me the difference between "a QDD and a servo" - buddy I've got bad news for you) -everyone picking a weird "product" name for their non-peer-reviewed-AI-generated-3D printer gripper projects. In academia, people used to title the paper usefully. Now we say: "MagicTact: a really small and (we promise) totally novel sensor where we put a magnet and a hall sensor into rubber and apply trigonometry" ...seriously? We're trying a bit too hard to rebrand the basics. You can name it when you invented it -- until then, stop trying so hard to have jargon. Nobody in the real world says half of this stuff out loud. Maybe this is my own failing, but when I hear it, I assume you are, at best, young, enthusiastic, and not very experienced. The whole point of faking it is to gain credibility and I think many people have missed that. Now get back into the pit and don't leave until people acting cute on the drawing feels like a dangerous choice. 💀 Only then will you be truly free. 🤖

Slick design
🧵After testing every hand glove on the market, here’s why we rebuilt ours around magnetic encoders: After seeing just about every hand data glove on the market, I’m convinced most of them are solving the wrong problem first. Here’s the thing nobody likes to say out loud: if your glove can’t tell you exactly where each finger joint is, every downstream model inherits that error. Hand data is only as good as the joint angles underneath it. So when we built DAS Dex 2.0, we started there. Why not soft gloves? They deform. The surface changes shape as you grip, so the contact geometry is never fully modeled, and mapping deformation back to a 21-point hand model introduces errors that need repeated recalibration. Comfortable to wear, hard to trust as ground truth. Why not EMF? Electromagnetic tracking drifts near metal and EM sources — exactly the environments manipulation data gets collected in. Why not pure IMU? It needs constant recalibration and drifts with temperature and time. So we went with magnetic encoders. An exoskeleton with a built-in skeletal model reads joint angles directly — no soft-body deformation to back out, no line-of-sight dependency, strong immunity to EM interference, and no constant recalibration. You read the joint, you get the angle. The classic objection to exoskeletons is bulk. That’s the real engineering story here: we built a custom miniature magnetic encoder just super thin and moved the whole structure to the side of the fingers instead of the back of the hand. 21 encoders per hand, 150g total (without wrist cam) — thin enough that it gets out of the way while you work. On the sensing side: a self-developed full-palm thin-film e-skin, over 980-point high-density array covering the whole palm (not just fingertips), reading normal force down to 0.1N. A 3D silicone surface adds grip. We tuned specifically for low cross-talk — press one point and it stays one point, no phantom spread. Put together: direct joint readout + full-palm tactile + wrist camera + IMU, in a glove light enough to forget you’re wearing. That’s what we think ground-truth-grade hand data collection should look like. We just debuted DAS Dex 2.0 at CoRL 2026. If you’re working on dexterous manipulation, teleop, or hand data pipelines and want the technical details — benchmarks, tactile specs, how we validate joint accuracy — my DMs are open. Genuinely love to meet everyone there!

Going to some recent deployment sites in the Bay Area for robots, I am seeing a lot of instances of a robot thinking it grabs something, and carrying the imaginary object to the end position. This is embarrassing. Using a learned policy for everything just converts easily detectable failures into needlessly difficult AI problems. Like...if you're not holding anything, whether that's determined visually or via a force sensor, can't you just reset and spam retries? Is this whole thing running inside of a global state machine or not? What fraction of today's VLA failures disappear when you add some basic physical state checks? Why do new-age software engineers want to black box everything? Why can't they be tasteful in their applications of frontier technology? Do they care about success, or novelty more? Our problem is not just the fact that robotics engineers are rebranded software engineers now. It's worse than that....they are rebranded software engineers who don't understand system architecture. I know you're working with off-the-shelf hardware and off-the-shelf models and you are fueled by matcha and SF delusion, but damn, it would take like one day to catch some of these bugs. Have some pride in your work. Get it working. Don't be a leech, contribute to the positive public opinion of robot usefulness.

To all the crypto VCs flocking to SF: please don't turn robotics into your next speculative bubble. This industry doesn't need inflated valuations, manufactured hype, and another boom-and-bust cycle. It needs patient capital, real engineering, and long-term conviction. Robotics founders need to be more selective than ever about whose money they take. The wrong capital can destroy an industry faster than the wrong technology.

Also he casually uses 'convex hull' in a metaphorical sense which is extreme green flag

Just had the pleasure of seeing Quan Vuong (co-founder at Physical Intelligence) talk about robotics at @fdotinc today. I saw him speak once before. I thought it was a fluke the first time, but I like this guy. He's sharp and futuristic, but also conservative and pragmatic. This is what it looks like when a researcher successfully transcends and becomes a leader. Nothing better than watching a domain expert say "I don't know". Throwing robots into molten steel is cool too, but this is where it's at.

Look at the contrast between Figure AI throwing their robots into molten steel vs the @UnitreeRobotics HX-2. It's not lost on anyone that this robot could also be used to cause harm, but this video is a very optimistic take. I've always thought it was weird that people couldn't see any applications in a highly athletic robot. Protecting someone from an attacker, saving a child from being hit by a car, etc. In the US, we're building geriatric little robots that can fold laundry and load the dishwasher. In China, they are trying to build superheroes. With manipulation control models rapidly becoming a race to the cost floor, to accuse Chinese robots of merely being parkour demonstrations is short-sighted. The difference is very simple… In the US, a robot won't succeed unless it can independently spur economic growth or make some people a lot of money. In China, the government can simply benchpress new sectors into existence through strategic partnerships and spending with private industry. I'm sure one of these things will be asking me to see my visa the next time I'm in China. Places like crowded city centers and airports will have them first. The robot cannot fear for its bodily harm, so the potential for it to be a protector (rather than an antagonist on a bad day, à la police officer) is there, depending on the choices of the creators.
Unitree Introducing: H2-X Kung Fu Legend Guardian🤖 Powerful performance to serve and guard you, with a maximum endurance of approximately 8 hours. Please use the robot safely and in a friendly manner.


We are open-sourcing V1.0 of our agentic robotics framework, ASPIRE, with full real robot skill discovery and self improvement capabilities! On the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night. And the frontier is already on the move: - Opus 5.5 slightly edges out Astra on our robot simulation bench. - Sonnet 5.5 hits the Pareto frontier of cost vs. robot capability, $12/task. We first introduced ASPIRE and /robot skill library in June, way before the big models could reliably do robotics. A rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks. Now you can also feel the Physical Bitter Lesson at home like we did on that fateful day ;) links in thread


Seems like every other white paper claims that their idea can be easily commercialized, usually written by a bunch of PhDs who have never shipped a product in their lives. And authors can claim novelty now because people are widely looking at robot arms again, so they take half-century old concepts and reapply them to whatever open-source little 3-D printed arm they found connected to whatever hugging face software setup was vibe coded for them by their Claude subscription. It's OK to be honest about when you are applying old ideas to a more receptive audience - you don't have to claim everything is a breakthrough.

SF Tech Week opened people’s eyes to just how much better Luma is than Partiful 😭

Who still thinks their models and data are a moat?
Let your app choose the right model, tool, or action in near real-time with Decisions API, now available to all developers in public beta. The Decisions API makes decisions up to 10x faster than GPT-6 Luna through the Responses API.