Physical Intelligence

Physical Intelligence is a San Francisco-based artificial intelligence company developing general-purpose foundation models for robotics. Its stated aim is to bring general-purpose AI into the physical world by creating learning algorithms and a single model capable of controlling any robot to perform any task, replacing the narrow, hand-engineered programming typical of industrial robots [0][1]. The company describes itself as a group of engineers, scientists, roboticists and company builders [0][1][5].

Its core products are a series of vision-language-action (VLA) models. π0 ("pi-zero"), released October 31, 2024 after roughly eight months of development, combines internet-scale vision-language pretraining with open-source robot datasets (including Open X Embodiment) and the company's own dexterous manipulation data collected across eight distinct robot platforms — UR5e, bimanual UR5e, Franka, bimanual Trossen, bimanual ARX, mobile Trossen and mobile Fibocom — to output low-level motor commands at up to 50 Hz from a 3-billion-parameter VLM backbone [3]. Subsequent releases include π0-FAST and an action tokenizer (FAST) reported to speed training 5x, π0.5 with open-world generalization for mobile manipulation, π*0.6 trained with reinforcement learning from experience, work on real-time action chunking, multi-scale embodied memory (MEM) for tasks longer than ten minutes, online RL for precise manipulation, and π0.7, described as a steerable model with emergent capabilities [0][1][2].

The company positions these models as a reusable "physical intelligence layer" analogous to language-model APIs, intended to lower the cost of building robotics applications across homes, hospitals, offices, warehouses and factories [4]. Reported end-use demonstrations include laundry folding, coffee/espresso preparation, table bussing, grocery bagging, cable routing, box assembly and e-commerce order packaging [3][4][6].

Founding story

Public sources state Physical Intelligence was founded in 2024 [5][6][7]. Reported founders include Karol Hausman (CEO, previously Google DeepMind), Sergey Levine (Chief Scientist, UC Berkeley professor), Chelsea Finn (Research Lead, Stanford professor), Brian Ichter (previously Google Research), Adnan Esmail, Lachy Groom (described as COO and previously a product leader at Stripe), and Quan Vuong [5][7]. Sources describe the founding premise as extending the foundation-model approach that succeeded for language to embodied systems, so that robots learn from diverse cross-robot experience rather than being reprogrammed per task [5]. The first model, π0, was built over about eight months and published in October 2024 [3][5].

Business model

The company develops robot foundation models and works with partner companies that deploy robots commercially, supplying models that those partners fine-tune and run on their own hardware and customer sites [4][6]. One description of the workflow cited by The Robot Report has developers streaming RGB-D camera images from any robot to Physical Intelligence's runtime, with the system tokenizing the visual stream and movement history for a 3-5 billion-parameter transformer and a hardware abstraction layer converting output tokens into robot-specific joint commands [6]. Some models, including π0 and π0-FAST, have been released open source [0][6][7]. Sources do not describe pricing or contract structures.

Traction

Partner-reported results indicate measurable deployment progress. Weave reported that π0.6 significantly increased autonomy versus π0.5 in laundry folding at a live San Francisco laundromat, and that including Weave data in pre-training reduced missed grasps by 42% and interventions by 50% per laundry load [4]. Ultra reported successive step-ups in intelligence, throughput and reliability from π0 to π0.5 to π0.6, including a full-shift order-packaging run at 96.4% autonomy and higher throughput when Ultra data was added to pre-training [4]. The company also reported that RECAP-trained models doubled throughput on tasks such as inserting an espresso filter, folding unseen laundry and assembling cardboard boxes [6]. A secondary profile cites π0.6 success rates exceeding 90% on such tasks [5].

Latest developments

On April 16, 2026 the company published π0.7, described as a steerable robotic foundation model exhibiting a step-change in generalization [0][1][2]. Preceding releases in 2026 covered efficient online RL via an extracted RL Token (March 19, 2026), Multi-Scale Embodied Memory (March 3, 2026), and a February 24, 2026 post outlining the "physical intelligence layer" concept with partner results from Weave and Ultra [0][2][4]. In March 2026, Bloomberg reporting relayed by Tech Funding News indicated the company was in talks to raise about $1 billion at a valuation above $11 billion, with Founders Fund and Lightspeed Venture Partners potentially participating and existing backers Thrive Capital and Lux Capital in discussions to return; the deal had not closed [7].