Physical Intelligence
What changed, and what is the evidence.
www.pi.website
Physical AIDeploymentsModelsComputeDataHumanoidsControlInferenceHandsSensingVisionHandsPolicyActuationCamerasFactoryFundingGpuCobotsEnd EffectorsRuntimeFccFleetCameraNpuQuadrupedsWorld Models
🗞️ Pulse first · 3 news posts
Coverage
Related topics
Timeline
What changed, and what is the evidence.
Oct 22, 2026 · Post · posted
| Headline |
|---|
A year ago, running one coding agent was the whole workflow. Now people are stringing together three or four just to ship a feature, like the builder who posted today about coordinating Claude Code, Codex and Pi through a separate tool… |
"Off-the-shelf AI code tends to be pathologically slow in the stupidest of ways." @castorhat on taking π0.5 inference from 230 ms to 87 ms. Here's each fix and what it bought: https://t.co/M906rOeXbn |
We cut π0.5 inference from 230 ms to 87 ms on a laptop GPU, without pruning or distillation. The rollout code from our IROS demo is open source: https://t.co/Syfh1SiQfJ https://t.co/vBzcG5BSsU |
A Python pipe buffer was holding camera frames until 64 KB had piled up. That quietly added hundreds of milliseconds to our π0.5 rollouts. New from @castorhat: how we found it, and what changed on the robot once it was gone.… |
Turns out that simply extracting policies with supervised learning conditioned on the advantage can get state of the art results in RL, as long as we get a few important details right! https://t.co/fZsaoIcUgs |
We nudged the cameras on a π0.5 policy a few degrees and a few centimeters at a time. Lifting barely changed. Placing fell apart. What @castorhat's fine-tuning ablations found about the most overlooked part of a robot rig:… |
Robotics is entering the API era. ★ Dream Labs | Dream-1 | From $3/hr ★ Google DeepMind | ER 2 | $1 in/$5 out ★ Generalist | Waitlists ★Physical Intelligence | Waitlists Robot manufacturers can focus on hardware instead of building their… |
We built a path-traced YAM simulator to run thousands of π0.5 rollouts a day without resetting a scene. It's open source now, so you can run the same ablations on your own rig: https://t.co/Gpf69r6Npi https://t.co/Ms8aXpxhuj |
Hard lighting cut our expert-only π0.5 checkpoint's success roughly in half. The same lights cost the full fine-tune nothing. New from @castorhat: the fine-tuning ablations behind our IROS demo. https://t.co/JciqcCBMYy |
Massive week in robotics. I went through everything from Google DeepMind, Physical Intelligence, NEURA Robotics and even more. Get your coffee and see what happened in robotics space: |
Your phone probably costs more than this humanoid robot. Zeroth 01 is an open source humanoid 40 cm tall with a parts list starting at $350. Most of the frame is 3D-printed Feetech servos drive the joints and you control it in Python. One… |
🚀 #Rockchip RK3576 + #RaspberryPi PoE+ HAT: Power Over Ethernet Meets Edge AI! We adapted the #reComputer #RK3576 Module to the Raspberry Pi #PoE+ HAT, enabling PoE-powered operation through Ethernet while running #YOLO11 from Seeed… |
We trained π0.5 for 8 hours on 4× H100, streaming a 10.7 TB dataset straight from the Hub. Time spent waiting for data: 0.52% of step time. No full download first. Just add --dataset.streaming=true How it works 👇 https://t.co/nIo1kcO1k8… |
This Week in Physical AI 🧵👇 What happened in robotics and embodied AI over the past seven days, featuring @alexkoch_ai, @physical_int, @GoogleDeepMind, @RhodaAI and more. Let’s go 🧵👇 |
Edge VLM placement is decided by energy per token. Running Qwen2-VL-2B, a Raspberry Pi 5 with a Hailo-10H NPU cuts energy per token about two-thirds versus CPU, and avoids thermal throttling. Jetson Orin Nano wins short replies; the curves… |
Physical Intelligence made fine-tuning a frontier robot model an API call. The loop they describe: collect demonstrations, train, test on the robot, collect again. The compute is handled. The demonstrations are still the input. We've been… |
🤖 Pick a task. Watch your robot arm get to work! SO-ARM + pi0.5 VLA supports 3 trained pick-and-place tasks. 👀 See → 🦾 Pick → 📦 Place No coding. No manual teaching. Just pick a task! 🔗1000 credits for free: https://t.co/k57gf78SXu… |
🚀 Deploy YOLO11 on RK3576 Module + Raspberry Pi CM4 IO Board!! We tested the #reComputer #Rockchip #RK3576 Module on a #RaspberryPi #CM4 IO Board and successfully deployed #YOLO11 from #reComputerAILab for real-time object detection! ✅… |
A strong speaker lineup, live robot demos, and lots of robotics folks coming together... this should be fun. https://t.co/uyWCS3yBOK |
Most robotics talks are about capability. This one is about character. Do Park (@alrightdo ), Head of Design at @lightberry, opens the Humanoid Hub Conference with "From Pixar to the Physical World: Designing the World's Most Interactive… |
Access robotics intelligence via API. A new paradigm https://t.co/SWwpr3rgj0 |
Speakers from 1X, Dyna Robotics, Skild AI, Google DeepMind, Physical Intelligence and more. Humanoid robots, dexterous hands and more cool robotics demos on the floor. The robotics community in one floor, and four NVIDIA Jetsons to give… |
Physical Intelligence just opened its API beta. You upload data, fine-tune their latest robot foundation models on PI-managed compute, and run inference through a hosted gateway. CloudChef demonstrates workflow results:… |
Most "general" robot models quietly settle for simple grippers. Nature Will went the hard way. Natura-1: a 50B mixture-of-experts model trained on 1 million hours of human video, driving 20-DoF five-finger dexterous hands at a 100ms… |
Lux family co @goformic + @physical_int https://t.co/zsE8mFTZZT |
SO SICK! Embodied intelligence for every species of robot. Whether you like humanoids or bizarre robots. As Darwin said, 'endless forms most beautiful' Lux family co @physical_int https://t.co/Pb0YVFymHz |
PI’s frontier models are powering everything from harvesting and household robots to assembly systems and wheel loaders. Formic’s demo looks really good. The robot is handling the full box-packing cycle with PI’s models. Looks like they’ve… |
great news the barrier to entry into robotics has just dropped significantly i can imagine a world where building a specialized robot will be much simpler in practice let's make more robotics companies :-) good job @hausman_k… |
Honestly quite exciting to see so many different types of robots using the same intelligence layer. Kudos to the PI team for making their models easy for the community to use and build upon 👏 https://t.co/XXHgGdjsZd |
Physical AI feels like it's moving into a new stage. Physical Intelligence just opened the Pi API beta. Robot companies can upload their own data, fine-tune Pi's general-purpose foundation models, and run cloud inference to control their… |
API access for robot intelligence https://t.co/WQ4dJ1wtoC |
Starting a robotics company should be much easier than it is today. We're opening up access to our models to help make that happen. https://t.co/Dnb01EAGmY |
The Pi API is officially in beta! 🤖 Many people have already been using it to fine-tune π models for their robots -- below are a few We’re hoping to expand testing significantly. Let us know if you are interested! Blog:… |
Even a year ago I don't think I would have believed that PI models can run on such a huge range of different robots. https://t.co/F4tc2g1mt2 |
Announcing the beta release of the Pi API. We’ve partnered with companies across diverse industries to adapt our models to their robots and applications, from harvesting strawberries to assembling servers. Now we’re opening access to more… |
Super-impressed: Nature Will releases 50-billion parameter general foundation model for robot intelligence The big question: Is the robotics industry wasting billions on training as the frontier models & this foundation model get better… |
Back in April, I posted about Nature Will @naturewill when they first caught my attention. Now we finally get to see what they’ve been building. Their first model Natura-1 is trained on nearly 1M hours of egocentric human data, combining… |
This might be one of the clearest tests of physical intelligence. A robot can learn a manipulation skill. The harder case is using that skill when the object environment or instruction changes. Nature Will @naturewill ’s Natura-1 is built… |
"Experiments in building products at Physical Intelligence" -- Quan Vuong's (@QuanVng) keynote at the Humanoid Hub Conference. He co-founded @physical_int, and before that helped build the foundational RT-2 and OpenVLA models at Google… |
At #electronicafair, we're showcasing the technology transforming data into intelligence and intelligence into action. Join us in Munich to explore how physical intelligence is enabling more perceptive and resilient systems across… |
Silicon Valley concentrates the venture capital and model talent. Manufacturing plants run on takt time, not demo schedules. Physical AI gets proven when an autonomous policy survives an active assembly cell. https://t.co/Hgj0lyHNHQ |
Open Duck Mini V2 is a 42 cm open-source robot duck with a parts budget under $400. Its walking controller is trained in MuJoCo and runs on a Raspberry Pi. I like that the team shares the CAD files and code so others can study how it… |
I am trying to visit them all and have only gotten a small percent. https://t.co/MCCp6V9nuE |
This is the capital of robotics! 🌁 Silicon Valley is home to so many physical AI companies that you could spend a month visiting them and still not see even 10% of them (trust me, I tried). So, the Bay Area is the world's leading… |
Lux family co @physical_int cofounder @chelseabfinn 🤖 https://t.co/gFJQzNkiIS |
🚀 Do you know reComputer RK3576 Module can work with Raspberry Pi CM4 IO Board We're testing compatibility between the #Rockchip #RK3576 Module and the #RaspberryPi CM4 IO Board, with several interfaces already up and running: ✅ HDMI… |
Training a robot to grasp an object means learning from many attempts, viewpoints and situations. One robot with three cameras at 20 fps produces 216,000 images per hour. Now that robotics is scaling, how can we efficiently train on all… |
Ultra has raised a $50M Series A, bringing its announced funding to $62M. Its warehouse robots pack orders using Physical Intelligence’s AI, with humans stepping in when needed. Ultra says they’ve already packed more than 500,000 orders.… |
A humanoid robot that can catch a moving cup Most robots freeze the second something starts moving. Cup on a belt, shifting grasp, anything not stuck in place --and they miss. Long-WAM from NVIDIA (with MIT, HKU, UCSD) went after that. On… |
What happens when #AI agents stop living on screens and start controlling cameras, wheels, monitors, and eventually full #robotic bodies? I joined Samson Rose’s Founder and host of the Machine Minds podcast, @Toroosian, to discuss the… |
Renting design is the easy part. Qualification cycles for safety-critical actuators run 14 to 18 months per configuration. Tooling validation, supplier audits, IEC 62443 compliance. That pipeline does not compress because you found a… |
Boston Dynamics didn't have a booth at IROS 2026. It still took the week. @castorhat's recap from Pittsburgh covers a show floor crowded with hands, hand makers turning into data companies, and the site-unseen demos that held up ⤵️… |
Field notes from IROS 2026 in Pittsburgh: a show floor full of dexterous hands, hand makers turning into data companies, and site-unseen demos that held up. https://t.co/76rYBC8l6B |
Cortex Harness feels like a Physical AI agent platform. It supports general-purpose models like GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.1... alongside robot-native VLA models like MolmoAct2 and pi0.5. TBH, this demo is really cool-- A… |
Google DeepMind, Boston Dynamics, Skild AI, Physical Intelligence… quite a lineup. A who’s who of physical AI and humanoid robotics at Encord’s Long Horizon summit. Use our code HUMANOIDS20 to get 20% off. Oct 20–21 · Fort Mason, SF… |
I can't remember the last time this many people from physical AI were booked into one room! The Long Horizon Physical AI Summit by @encord_team lineup is fire! 🔥 → Sebastian Thrun, founder of Google X, who led the self-driving project… |
Pi 0.8 announcement 👀 https://t.co/Ziyb45pVWG |
Why will Boston Dynamics win the humanoid robotics race? We've already commercialized autonomous mobile robots, creating markets with Spot and Stretch. Now we're doing it again with Atlas: learning in real environments, designing for… |
Here's a new home robot for $3,500: Flourish 1. Teach it tasks with your phone. Or get skills from an App Store. - Raspberry Pi chip - Super-basic grippers - 12-hr battery - Wheeled base - 1.5 kg per arm Can pick up stuff, water plants,… |
We are excited to announce that World Labs is joining @AMD. The research and technical breakthroughs we have achieved since our founding in 2024 have given us a clear vision for AI’s potential to solve problems in the spatial and physical… |
Humanoid robots are improving fast, but flashy demos still hide the harder question: how many are actually working reliably in the real world? @Rewkang and @GoingBallistic5 sat down with @herbertong to separate the hype from deployment.… |
What is #physicalAI? Our CTO Pras Velagapudi joined @MIT_CSAIL to dive deep into physical intelligence. Worth a watch: https://t.co/ykzUoF1XyV |
Coachella of AI is here. Still some tickets available at https://t.co/ODnW6z1Ff5. Line up is ready: Adarsh Kulkarni, Foundry Robotics Alexander Toshev, Wayve Anastasis Germanidis, Runway Andy Chen, Runway Anthony Maggio, Runway Ashley… |
Midcentury's stealth launch for robotics infrastructure is a critical unlock for physical AI. Scaling LLMs relied on massive internet-scale datasets and evaluation. Robotics has lacked that. Building the world's largest egocentric dataset… |
IRON is starting to look less like a machine you operate and more like a machine you interact with. @XPENG_Global is testing four abilities that matter for real-world interaction: recognizing people, remembering identities, switching… |
Imitation learning, especially with interventions, has driven so much recent robotics progress. However, improving a policy via targeted interventions until it reaches a useful and deployable success rate is a time and labor intensive… |
Humanoid robotics may be approaching its ChatGPT moment. @Figure’s Helix 2.5 transferred learned behaviors across 30 unseen homes. The bigger story isn’t bed-making or laundry. It’s whether physical intelligence can start scaling with data… |
Researchers have trained a robotic hand to walk on its fingertips without an arm. Using an off-the-shelf WUJI hand equipped with a Raspberry Pi, IMU, and battery, they trained reinforcement-learning policies in NVIDIA Isaac Lab. The… |
Thing from The Addams Family has a robotic cousin. @amrhkzp from ETH’s Soft Robotics Lab taught an off-the-shelf WUJI hand to walk on its fingertips, with a Raspberry Pi and battery on its back. It can also recover, interact with a… |
An Anthropic acquisition of Physical Intelligence would validate the deep shift towards embodied AI. For automotive, the impact isn't just in manufacturing but in how autonomous systems perceive and interact with the physical world, moving… |
Legacy microcontroller architectures cannot support real-time transformer inference at the joint level. Major analog players acquiring embedded AI silicon startups confirms that physical intelligence is moving directly into localized motor… |
With the ICRA deadline only a few days away, it's a great time to look over my guide for how to fake a robotics paper: - Don’t let other people run comparisons. If other people can try your model out, they’ll quickly uncover its… |
The pattern I keep watching across AI infrastructure: software standards always arrive before the hardware catches up. MCP gave AI agents a common interface for software tools. Every model, every framework, one protocol. It worked.… |
NVIDIA's competitive advantage has never really been about chips alone. It's been about the software moat - CUDA, cuDNN, TensorRT, the full stack that makes their hardware irreplaceable in production ML workloads. SoL-Pi is an interesting… |
Analog Devices to Acquire Alif Semiconductor, Adding an AI-Native Processing Platform to Advance Physical Intelligence for the Next Generation of Real-World Systems. https://t.co/KyJo4B9EfK https://t.co/SM0Po9UgbO |
This might be one of the clearest snapshots yet of how crowded the humanoid robot race has become. 16 humanoid robots ➝ AGIBOT A2 ➝ EngineAI T800 ➝ Xiaomi CyberOne ➝ XPENG IRON ➝ Figure 03 ➝ Galbot ET1 ➝ Unitree humanoids ➝ Boston Dynamics… |
Last Week in Robotics (09.01-09.07) 🤖 Traditional Robotics • @Figure_robot signed a compute deal with @nscale for up to 100K NVIDIA Vera Rubin GPUs, starting with a $3.5B commitment to scale Helix training • @xdofai entered late-stage… |
Galactic scale narratives ignore terrestrial manufacturing constraints. Scaling physical intelligence requires precision reducers, magnetic materials, and high-voltage power electronics. The near-term bottleneck is precision machining… |
Edward Luo built this himself: an app that runs Claude Code, Codex, Cursor, Grok, Antigravity, OpenCode, and Pi side by side in the Mac notch. Agent finishes, notch pops. Needs approval, notch pops. One click back to that session. The… |
# openpi openpi holds open-source models and packages for robotics, published by the [Physical Intelligence team](https://www.physicalintelligence.company/). Currently, this repo contains three types of models: - the [π₀… |
π0: A Vision-Language-Action Flow Model for General Robot Control Physical Intelligence Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon… |
Hosted By Edmond Jono Hart James (Jingxi) Xu Mene Mazarakis Angela Zhang Manfredi Bernardi Adam 134 Went Akshobhya Gupta, Hannah Tsui and 132 others Contact the Host Report Event AI 🤖🍨 Sundae Robotics 06: V-JEPA 2 & Predicting Physical… |
@Arifx0001 @BitRobotNetwork Great synergy leveraging high-end and affordable tools for teleop to scale AI efficiently |
Stats time! Humanoid robot startups raised $8.6B this year. Almost none of them ship robots. The company shipping 44% of the world's humanoids is worth a fraction of the one still running pilots. Here's the humanoid money board, valuation… |
Warehouse China Warehouse US Warehouse Germany Warehouse The store will not work correctly in the case when cookies are disabled. US Warehouse: Enjoy FREE UNIUNI shipping on orders under $50! (Excludes XIAO & Raspberry Pi series) \\ \\ the… |
Post Log in Sign up Post Timothy B. Lee on X: "Last year roboticist Benjie Holson created the "Humanoid Olympics," a list of tasks that are trivial for humans but hard for robots. A startup called Physical Intelligence solved several of… |
Skip to content You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your… |
i thought community robot data would hit diminishing returns fast but Axis Dataset V1 just made that take look stupid @axisrobotics pushed π0.5 from 83.9% to 88.8% on LIBERO Plus, and performance kept climbing as training data scaled from… |
Skip to main content Search arXiv Press Enter to search · Advanced search Computer Science > Machine Learning arXiv:2604.15483 (cs) \Submitted on 16 Apr 2026 ( [v1), last revised 24 Apr 2026 (this version, v2)\] Title:π0.7: a Steerable… |
🤖 Something is changing in Physical AI. For years, humanoid robots were mostly about impressive demos. This week feels different. $900M+ raised by XPeng’s robotics business. Figure committing over $1B to data and compute. Hyundai… |
AI is still so early, but so bullish. We're witnessing a huge unlock of economic value with the adoption of AI models like Claude, and more recently, agents like Grok Bot. This then filters down into the most bullish of bull points for AI.… |
Fifteen years after "software is eating the world," Marc Andreessen's own firm just admitted atoms are eating software's margins. SUMMARY This is not an a16z story. a16z's $1.1B Machine Age Fund is simply the loudest confirmation yet of a… |
서랍이 안 열리자 수저를 오븐에 넣었다. Physical Intelligence를 만든 Sergey Levine이 Ryan Peterman 앞에서 말한 평가 장면이다. 모델은 π0.5다. 주방에서 수저를 넣으라고 했는데 서랍이 안 열렸다. 옆에 오븐이 있으니 그걸 열고 넣기 시작했다. 아이한테 치우라고 하면, 안 보이게만 넣기도 한다는 것이다. 접시 닦기도 그랬다. 회색, 초록, 하나를 떨어뜨리자 본체가 그 위로 가서… |
In robot control ER 2 competes with NVIDIA's open Isaac GR00T family (GR00T N1.7, Cosmos Reason), Physical Intelligence's π-series, Figure's Helix, and OpenVLA. In interoperability standards, MHS enters a space already occupied by SiLA 2… |
Many People believe Physical might be unachievable not because robots lack arms or GPUs rather, because robots do not have enough real training data. Normal AI LLM can train on text that already exists on the internet but Robots cannot. A… |
Physical Intelligence is bringing general-purpose AI into the physical world. We are a group of engineers, scientists, roboticists, and company builders developing foundation models and learning algorithms to power the robots of today and… |
@ClementDelangue While SoftBank pays $6B for humanoids that haven't shipped, Hugging Face ships a $399 duck you can actually train tonight. The Raspberry Pi playbook applied to robotics — the cheap hackable one is how a generation of… |
Post Log in Sign up Post Rhys on X: "Cool to see a bunch of new robotics foundation model releases: π0.7 from @physical\int GEN-1.5 from @generalistAI S1 from @SkildAI Feels like the GPT-3 moment for robotics. Robot Models are Few-Shot… |
scaling laws do not apply to physical AI yet. i'll say it again. scaling laws do not apply to physical AI. chelsea finn from physical intelligence showed this during her YC talk. they trained pi0.7 to make coffee. the robot couldn't learn… |
$14B valuation. Pre-revenue. A handful of customers. That is Skild AI today. And it is not the outlier. Physical Intelligence: $11B Generalist: $2B to $3B in months, $600M raised Skild AI: $14B Three companies. Billions in capital. Almost… |