


Physical AI & Robotics◽️Growth & GTM Engineer @tnkrdotai ◽️Head of GTM @DiracRobotics ◽️GTM @Hyphenbox


Introducing DrivingBench v2 🧵 We wired a Tesla Model 3 so any AI model can drive it, and put 8 frontier VLMs (Claude, GPT, Grok, Gemini models) and 6 fine-tuned robot models (VLAs & WAMs) behind the wheel. A new lot and two new courses, none of which they’d seen. Only one VLM ever reached the finish, and its fastest run took 26 minutes. 🏁

Here a @Universal_Robot (UR5) arm threads a cable through a car door, running fully autonomously. From collecting demonstrations to training and deployment, the entire process was completed on Neuracore.

Physical Skills Lamp is a physical AI agent. And Lamp's got physical skills. Dance. Security Guard. Drink Water Reminder. And many more. autonomous.ai/lamp This week, we're focused on shipping the Skill Store (aka the App Store for Lamp). So you can easily install a new physical skill for Lamp. Another design goal is to make it super easy for you to prompt a new physical skill. As easy as prompting the agent skills you already know. You can keep the skill for yourself or you can publish it for others to use.


Our in-house, human-scale humanoid is now walking around our office with our own policy :) Lots more to come! youtube.com/watch?v=LebJCFC_…

A one-sided conversation. You can have a conversation silently with Subvocal (YC F26). I couldn’t hear Arnav, but the glasses could.

Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

A modified SO-100 running UMI policies and manipulating objects outside its training examples. Pretty remarkable what affordable robot hardware is starting to support.
Today I'm open-sourcing Solarian: a Python framework to build, simulate, train and deploy AI robots. Plus the hardware to run Stanford's UMI policies on a modified SO-100 arm with a ~$100 gripper, instead of a $30k industrial arm and a ~$4k gripper. 🧵

The ambition isn't just to build better robots anymore. It's to make physical labor programmable at scale. $500M says investors are taking that possibility seriously.
Genesis is entering its next phase. Today we've raised a $500M Series A at $3.5B. We can imagine far more than we can build physically today. Almost everything we are creating is still paid for in human time, by human hands. We want to remove the physical limits on human ambition. We are building general-purpose robots with human level intelligence and capability. Ones that work reliably across entire workflows for years. There are countless versions of possible futures where billions of robots are deployed everywhere, and we aspire to bring the brightest one that carries our hopes into reality.

🚨 BREAKING: Genesis AI (@gs_ai_) has raised a $500M Series A led by Khosla Ventures, Premji Invest and Eclipse! That is up from a $105M seed in July 2025. @zhou_xian_ and @theo_gervet met as grad students in a CMU lab and founded the company in 2024. Xian says Genesis has “cracked dexterous manipulation.” Their strategy is vertical integration. Genesis generates its own training data, builds its own model and designs the humanoid alongside both. → Force data comes from sensor-enabled gloves, worn along with headset cameras by 100+ real workers around the world → The robots wear the same gloves, and their hands copy human proportions → Millions of high-fidelity simulated experiences come from a physics algorithm co-authored by adviser Minchen Li → Flagship humanoid Eno: > 30 built so far, piloting with a handful of undisclosed customers → A US mass-manufacturing contractor is chosen to start next year, and the production design is finalized Eno looks nothing like a metal skeleton. Think “an iMac on wheels.” :D Xian's view on models is the part I find most interesting. He says architecture, whether VLA or world-action, contributes maybe 10% of the final result. What matters is the training infrastructure, the data infrastructure and data quality. “I don’t think we expect or we need another fundamental scientific leap. We have all the recipes. It’s just about execution.” That is a strong claim, and it will be tested on the factory floor. $500M buys a lot of execution. ROBOTICS IS HOT. 🔗 Read more here: techcrunch.com/2026/11/17/ge… ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com

The next wave of AI training data might come from millions of people simply documenting the world around them.
Introducing roots by micro1. The future of intelligence will learn from the real world. We’re making it easier for anyone to be a part of that process and earn from it. roots is an app that lets you contribute real-world data to help train AI systems, from photos and videos to screen recordings, wearable data, and more. Privacy and security are a priority as we build roots. This includes removing sensitive personal information from submissions, requiring consent when recording others, and securely storing data. Explore available projects, capture or upload your contributions directly through the app, and get paid for approved work. New opportunities will be added regularly. Join the waitlist via link in comments.

This is a fascinating approach. Instead of throwing a massive model at every search, distribute the work across smaller agents and coordinate the results.
Introducing SID-2, a new search model built on compute-scaling swarms. Our swarms find better results than Opus and Astra, while being 8x faster and 20x lower cost. By 2030, AI will write an internet's worth of text every 24 hours. Current search pipelines were built on hand-designed rules, adding more won't scale to meet this incoming volume of data. SID-2 is available on our API today. Try it on your most demanding searches.

We're watching the earliest stages of a completely new way for humans to interact with machines.
Introducing AIR-O, a model that decodes what a person is attending to directly from brain scans. Using only noninvasive brain data, we can generate an image from scratch in real time that depicts where a person is focusing their attention. This is the first of a new generation of algorithms that can read the shape of thoughts directly from brain activity.

Today we open-source the Robotics for @antigravity plugin, one of the best tools available for roboticists to go from 0 to execution, powered by Gemini Hydrogen @GoogleDeepMind . The plugin is a set of skills you can install on Antigravity. It allows to do anything you can imagine with robots: from asking it to install and calibrate new arms, to creating complex simulations and robot behaviours zero-shot In this demo, we asked it to create from scratch a program that can play chess with a Franka Duo. It created the sim environment, the control and calibration code, and finally translated it to the real robot. It was all created by Hydrogen, that even plays the game live, thinking of all the moves. More examples and link in thread

NVIDIA just dropped 10,000 OpenUSD objects on Hugging Face 🤯 → a microwave with a hinged door → a dresser with sliding drawers → a forklift with a lifting mast → a wheelchair with spinning wheels 1,000 categories, 8,222 with joints, real mass and collision. Ready for Isaac Sim, Apache 2.0. ⬇️ huggingface.co/datasets/nvid…