The article argues that, limited by onboard compute and heat dissipation in robot bodies, embodied model inference is shifting to the cloud, and embodied AI MaaS is becoming a new direction. In July 2026, Yuanli Lingji released a general embodied MaaS service; in October, Physical Intelligence launched an invitation-only Pi API Beta; and Dream Labs launched Dream API backed by a $90 million seed round.
After GPT-6 Astra's release, general large models can directly control robots to perform tasks; embodied data orders are rebounding, the value of body companies is being reassessed, and model companies building embodied brains now face competition on the same field as foundation model giants. In NVIDIA testing, Astra achieved a 66.3% success rate in zero-shot direct action output, higher than π0.5, but average single-inference time was 18.56 seconds, and one successful action consumes about 160,000 tokens, equivalent to $1.96. The industry currently prefers having general models handle planning and the semantic layer, while embodied models handle millisecond-level motion control and the execution layer.
When running reinforcement learning with prior data, a behavior constraint is important for trading off exploration against staying close to known good parts of the data. But choosing this constraint is delicate—we neither want to merely match all prior data nor to deviate too far. In our new paper, we propose a new constraint that combines a support constraint with a continuous penalty to achieve excellent offline-to-online reinforcement learning performance.
Foreign media report that Physical Intelligence is in talks for a new funding round, planning to raise at least $1 billion at a target valuation of about $20 billion, more than 8 times its $2.4 billion valuation in November 2024. The company was founded in 2024, develops the π series of robot models, and opened a beta version of the Pi API on October 13. The article also compares Skild AI: 10 months of formal commercial deployment, more than 60 paying customers, ARR exceeding $100 million, and cumulative recognized revenue of $50 million.
GXO announced on October 26 that Agility's Digit will handle e-commerce fulfillment at warehouses in Statesville, North Carolina, and Lebanon, Tennessee, starting in early 2027, beginning with Digit 4 before moving to Digit 5; at a Georgia site, Digit has already moved more than 100,000 totes.
In the 50 days since GPT-6 Astra was released, the embodied AI industry has conducted intensive evaluations and discussions around it, and general-purpose large models have begun directly taking over robot operation. Tests by SenseTime and several institutions concluded that Astra is basically capable of completing embodied tasks, but still has a clear gap from practical use in real environments, underperforming in millimeter-level contact, dynamic control, and response speed. The article also reviews industry uses such as using Astra directly for policies, reviewing VLA actions, building simulation environments, and data annotation, as well as chain reactions including cooling financing and major companies stepping up data collection.
Why it matters: It reviews evaluation results, company reactions, and industrial chain changes in the embodied AI industry within 50 days of GPT-6 Astra's release, presenting the differentiated landscape after general-purpose large models entered robot operation.
US robotics AI company Physical Intelligence is in talks for a new funding round of at least $1 billion, with a post-money valuation expected at about $20 billion, a sharp rise from its $5.6 billion valuation last November. The company has no commercial revenue to date and has not mass-produced its own robots; its core focus is general-purpose robot foundation models that adapt to third-party hardware. It has iterated from π0 to π0.7 and this month launched a beta API for its robot foundation model, offering model fine-tuning and inference services.
Enact Intelligence, a physical AI company headquartered in Zurich, Switzerland, has closed a €26.68 million (about $30 million) pre-seed funding round, with investors including redalpine, Notion Capital, and firstminute Capital.
On October 13, Physical Intelligence (PI) announced the launch of the Pi API beta, packaging robot data upload, hosted fine-tuning, cloud inference and policy generation, and evaluation feedback into a single service chain, in an attempt to sell embodied brains via API.
It turns out that getting a few important details right is all it takes: extracting policies using only advantage-based supervised learning can achieve state-of-the-art results in reinforcement learning (RL)!
The article analyzes how, with the releases of GPT-6 Astra, Gemini 4 Argon, and the Pi API stacking up, the semantic, simulation, and fine-tuning layers of embodied intelligence are being taken over by general foundation model companies, and the world model is being demoted from overall commander to physical verifier.
Speakers come from companies including 1X, Dyna Robotics, Skild AI, Google DeepMind, and Physical Intelligence.
On-site demonstrations will feature humanoid robots, dexterous hands, and other robots.
The robotics community gathers together, with four NVIDIA Jetson units to be given away free.
Agenda:
Physical Intelligence announced the opening of the Pi API beta, allowing developers to upload data, fine-tune the latest robot foundation model on its managed compute, and run inference through a managed gateway. The company has partnered with companies across several industries to adapt the model to their own robots and applications, with use cases ranging from strawberry picking to server assembly; it is now opening access to more teams, and CloudChef demonstrated workflow results.
Pi API has officially entered public beta! 🤖
Many people have already started using it to fine-tune π models for their own robots—here are some examples.
We hope to significantly expand the scale of testing. If you're interested, please let us know!
Blog: https://pi.website/blog/api-beta
Signup: https://pi.website/waitlist
Physical Intelligence announced that the Pi API has entered beta and is opening access to more teams. The company said it has worked with companies across multiple industries to adapt its models to their robots and applications, covering scenarios ranging from strawberry picking to server assembly.
"Experiments in building products at Physical Intelligence" — a keynote by Quan Vuong (@QuanVng) at the Humanoid Hub Conference.
He co-founded @physical_int and previously helped build the RT-2 and OpenVLA foundation models at Google DeepMind.
He will give us an inside look at the work of turning research into real-world value.
October 15-16, San Francisco
Physical Intelligence has opened the Pi API beta, allowing developers to upload data, fine-tune π models on the company-managed compute and infrastructure, and then run inference through a hosted gateway; the current interface already offers more data-efficient fine-tuning and improvement capabilities.
Why it matters: Pi API brings data upload, fine-tuning, and hosted inference into one ready-made pipeline, helping readers weigh building their own machine learning infrastructure against calling the API directly.
Ultra announced the completion of $62 million in funding, of which the $50 million Series A was led by @hiFramework; partner Physical Intelligence reposted congratulations. In the quote, Ultra argues that deployment is the bottleneck in the robotics industry: making robots perform well in controlled environments is relatively easy, but making them enter the open world, remain useful, and earn the trust of those around them is far harder, and this funding is meant to solve the distribution problem. Physical Intelligence added that combining general-purpose robot foundation models with real-world deployment is making such things possible.
A third-party technical report, "GPT-6 Astra as an Embodied Policy," co-authored by Galbot researchers, was open-sourced less than two weeks after GPT-6 Astra's release, testing its performance as an embodied policy in the RoboDojo and RoboLab simulation environments.
Chelsea Finn shared a video of a Pi robot deployed at Dandelion Chocolate, running fully autonomously with no human intervention. In her review, she said the hardest part was unexpectedly reliable stacking rather than box folding itself: the original table-mounted robot struggled to see the stacking state without a separately added camera, and each box was placed in a different position, with a misplacement causing the whole stack to collapse later.
Physical Intelligence co-founder Chelsea Finn spoke at Startup School 2026 about building general-purpose robots in the real world: reinforcement learning doubled robot throughput, and the system can run autonomously for hours. She believes robots are entering the "GPT era"—moving from specialized models to general systems that work across tasks and across robots. Sergey Levine remarked that Chelsea's talk was once again rock-star level, and that robots really can fold laundry and make espresso!
Although we share a lot of research work at @physical_int, we rarely talk about what it's like to work here. Here is a great blog post by @stone_tao about what it's like to work at Pi, and a bit about Pi's simulation!
At the 2026 WAIC Embodied AI Forum, six guests predicted that robots' ChatGPT moment is two to five years away, with Yao Maoqing saying two years and Xu Danfei saying five, though no consensus has formed on the criteria for judging it. The discussion focused on three main threads: the data flywheel must be built through real deployments rather than data collection alone; VLA and World Action Models (WAM) are not opposites, with PI continuing the VLA path through π0.7 while Dyna turns to WAM; full-stack capability is becoming a trend, but a full-stack business model is not necessarily the only answer.
Tomorrow I will give a talk at ICML on emergent physical generalization, including π0.7 🤖
3:15 PM @ SCALE workshop, Ballroom 201
https://scale-icml-2026.github.io/
Ars Technica interviewed several robotics researchers and founders, outlining how AI is driving the development of general-purpose autonomous robots and the data, safety, and reliability obstacles they face.
Physical Intelligence is developing a general-purpose control brain that can work across many kinds of robots, using a vision-language-action (VLA) model to teach robots tasks such as making coffee, folding clothes, and peeling vegetables, rather than focusing on a single machine as Tesla, Boston Dynamics, or Amazon do.
Physical Intelligence announced that its latest model, π0.7, has several emergent capabilities: it can control a new robot to fold shirts without any shirt-folding data, can learn to use appliances through language guidance, and can complete multiple dexterous manipulation tasks within a single model.
Physical Intelligence has released the general-purpose robot model π0.7, which it says can perform a variety of dexterous manipulations with speed and robustness comparable to fine-tuned specialized models, and shows preliminary signs of compositional generalization on tasks not seen in its training data.
Why it matters: The article uses two specific experiments, an air fryer and folding clothes with a UR5e, to show the actual performance of compositional generalization and cross-embodiment transfer, making it easier to judge the gap between general-purpose and specialized models.
Physical Intelligence has added an RL token output to its latest model, π-0.6, which is learned rapidly through reinforcement learning by a very small actor and critic, allowing the model to be fine-tuned for precise tasks in hours or even minutes without training the entire model.
Physical Intelligence proposes a new method called RL tokens (RLT), in which a vision-language-action model (VLA) outputs a special token to interface with a lightweight online reinforcement learning policy, improving the speed and success rate of fine manipulation phases with only a small amount of real data.
Why it matters: The original clearly explains the interface design between the VLA and lightweight online reinforcement learning and provides speedup data on four tasks, so readers can assess whether this type of method applies to their own fine manipulation tasks.
We developed a memory system for our models that provides both short-term visual memory and long-term semantic memory.
Our approach enables us to train robots to perform long and complex tasks, such as cleaning a kitchen or preparing a grilled cheese sandwich from scratch 👇
Physical Intelligence proposed a method called Multi-scale Embodied Memory (MEM) that provides vision-language-action models (VLAs) with both short-term and long-term memory, enabling the model to complete tasks lasting up to 15 minutes, such as cleaning an entire kitchen or making a grilled cheese sandwich from scratch.
Why it matters: The article identifies memory mechanisms as the bottleneck for long-horizon tasks and presents a dual-scale design that stores video for the short term and language for the long term, an approach that can be transferred to other work.
General-purpose AI models underpin some of the most exciting applications we now can't live without. We envision that an analogous "physical intelligence layer," built with models like π0.6, will similarly drive a new wave of applications in the physical world.
We have recently begun working with a small number of companies whose robots are already performing practically useful tasks in the real world.
https://www.pi.website/blog/partner/?v=1
In a blog post, Physical Intelligence argues that robots need a reusable physical intelligence layer, and it published real-world deployment data from partners Weave and Ultra using π 0.6.
Why it matters: The two deployers provided comparative data on autonomy rates, missed grasps, and throughput, making it easier to judge the practical gains that foundation model iterations bring in real-world settings.
We got robots to wash pans, clean windows, make peanut butter sandwiches, and more tasks!
Fine-tuning our latest model makes all these tasks possible, with interesting implications for robotics, the Moravec paradox, and the future of large models in embodied intelligence.
See details below!
By fine-tuning its latest π 0.6 model, Physical Intelligence achieved preliminary solutions for the gold-medal tasks in 3 of the five everyday manipulation challenges in the Robot Olympics proposed by Benjie Holson, and reached silver-medal level in the remaining 2, with an average success rate of 52% and task progress of 72%.
Why it matters: Using the same set of externally proposed everyday manipulation tasks to compare performance with and without robot pretraining lets readers judge the practical value of foundation-model pretraining for manipulation skills.
We found an emergent property of VLAs such as π0/π0.5/π0.6: as pre-training scale increases, the model learns to align human videos with robot data. This gives us a simple way to use human videos. Once π0.5 has learned to control a robot, it can naturally learn from human videos.
Physical Intelligence research found that in generalization scenarios where only human demonstrations appear, fine-tuning π0.5 with human first-person videos mixed into robot data alone improved policy performance by about 2x, without using any explicit transfer learning mechanism.
Why it matters: Experiments show that merely mixing human videos into fine-tuning improves the policy by about 2x on scenarios with purely human demonstrations, and this capability emerges as pretraining robot data scale grows.