What Physical Intelligence does

Physical Intelligence is a San Francisco-based AI company founded in March 2024 that develops Vision-Language-Action (VLA) foundation models intended to serve as a general-purpose "brain" for robots across embodiments and tasks. Its core technology stack is built on a 5B-parameter vision-language model backbone augmented with a 300M-parameter action expert trained via flow matching, with FAST-discretized tokens used for autoregressive variants. The model family has progressed from π0 (Oct 2024) through π0-FAST, π0.5 (open-world generalization), π*0.6 (reinforcement learning from experience), MEM (multi-scale embodied memory), RLT (online RL for precise manipulation), and π0.7 (compositional generalization), each accompanied by a long-form technical publication on the company's research blog.

The company serves three customer types: (1) robotics deployment partners that integrate PI models into commercial robots — including Weave Robotics (laundromat folding), Ultra (warehouse e-commerce packaging), Robot.com/R-noid (food/hospitality delivery), Teleexistence (Seven-Eleven Japan convenience stores), and AgiBot (China humanoid ecosystem); (2) academic and industry researchers via the openpi open-source repository on GitHub and Hugging Face; and (3) government and scientific institutions (Argonne National Laboratory's RoSA project for lab automation). The firm has no public API, no disclosed pricing or licensing terms, and no commercial product on the open market as of early 2026.

The business model is pre-revenue and research-first: operations are funded entirely by venture capital (~$1B raised across two announced rounds at a $5.6B post-money, with a third round at $11B+ in early discussion), and the emerging go-to-market relies on partner-led commercial deployments rather than direct sales of a licensed product. Capital intensity, talent pedigree (Google DeepMind, Stanford, UC Berkeley), and a broad proprietary research IP portfolio are the primary strategic assets. Multiple partners have published quantitative autonomy and throughput gains (Ultra 96.4% autonomy over a full shift; Weave 42% reduction in missed grasps, 50% reduction in interventions vs the prior model), but no revenue, pricing model, or commercialization timeline has been disclosed.

Name

Physical Intelligence

Legal name

Physical Intelligence

Website

https://pi.website

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

101–250 employees

Short description

Physical Intelligence is a San Francisco AI company founded in 2024 that develops Vision-Language-Action (VLA) foundation models as a general-purpose control layer for robots. It serves robotics deployment partners (Weave, Ultra, Robot.com, Teleexistence, AgiBot), academic researchers via open-source releases, and government labs.

Ownership category

akta.pro rank

Product category

Robotics Foundation Models

NAICS

Navigational, Measuring, Electromedical, and Control Instruments Manufacturing (33451)

SIC

Services-Computer Integrated Systems Design (7373)

akta.pro primary industry

Robot Software, Autonomy Stacks & Simulation (ROS, digital twins) (HDAAAIAK)

akta.pro secondary industry

Humanoid & General-Purpose Robots (HDAAAIAJ)

Keywords

• Robot foundation models
• Vision language action
• Embodied AI software
• Robotic manipulation AI
• Robot policy models

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Technology or R&D, Personnel, Infrastructure, Operations, Supply Chain

Revenue model

1. Pre-revenue / research-first model: As of early 2026, Physical Intelligence has no commercialized revenue and no publicly disclosed pricing or licensing model; the company is funded entirely by venture capital and pursues a research-first strategy of building general-purpose foundation models before commercialization.
2. Partner / licensing-style deployment model: PI collaborates with deployment partners (e.g., Weave Robotics for laundry folding at laundromats, Ultra for e-commerce order packaging) to integrate PI models into commercial robots; partner-published results indicate real-world autonomous operation and quantifiable throughput improvements, implying an emerging B2B model where PI provides the "intelligence layer" that partners productize.

Go-to-market motion4 records

Distribution channels4 records

Marketing channels7 records

Core offering

Physical Intelligence develops and ships general-purpose Vision-Language-Action (VLA) foundation models that let any robot perform any physical task from natural-language instructions. Its model family (π0, π0-FAST, π0.5, π*0.6, π0.7) is delivered to robotics deployment partners and integrators such as Weave Robotics, Ultra, Robot.com, Teleexistence, and AgiBot, with select base models and the openpi fine-tuning codebase released openly on GitHub and Hugging Face.

Differentiator

Problem solved

Functional benefit

Brands

• π (Pi): The company's stylized brand symbol, used as shorthand for Physical Intelligence on its website and in product names (e.g., π0, π0.5, π*0.6, π0.7).
• openpi

Products and services

• π0 (pi0) Physical Intelligence's first generalist robot policy / Vision-Language-Action (VLA) foundation model that combines web-scale vision-language pretraining with cross-embodiment robot data to power general-purpose robot behaviors; open-sourced in February 2025 with checkpoints for ALOHA, DROID, and LIBERO.
• FAST Action Tokenizer A general-purpose robot action tokenizer using discrete cosine transform plus byte pair encoding that compresses action chunks ~10x versus prior binning methods, enabling 5x faster training of dexterous VLA policies; released as a standalone open-source artifact on Hugging Face.
• π0-FAST A tokenized variant of π0 that uses the FAST action tokenizer, enabling smoother, faster, and higher-frequency robot control; achieved state-of-the-art on the LIBERO benchmark and is trained jointly on the Open X-Embodiment dataset and Physical Intelligence's internal data.
• openpi open-source release Open-source experimental repository on GitHub releasing π0 and π0-FAST code, weights, and fine-tuning recipes plus ALOHA, DROID, and LIBERO checkpoints so that researchers and integrators can adapt PI's models to their own robot platforms.
• π0.5 (pi0.5) An upgraded VLA with open-world generalization, trained jointly on web data and robot data so the model can follow language instructions in entirely new environments (e.g., homes the robot has never visited).
• Hi Robot A hierarchical system that combines a high-level VLA policy with a low-level VLA policy, enabling users to instruct a robot in stages and interrupt it to give new instructions rather than only end-to-end; delivers 40%+ higher instruction-following accuracy than GPT-4o high-level baselines.
• RTC (Real-Time Chunking) A diffusion/flow-inpainting-based inference-time algorithm that reconciles overlapping action chunks to enable asynchronous real-time execution of VLA policies on latency-prone robots without requiring training-time changes.
• π*0.6 (pi-star-0.6) π0.6 refined with Recap (Reinforcement Learning from Co-Assembly of Productions), a new RL method that enables the model to improve with experience and autonomously self-correct, reaching practically useful throughput on real-world tasks such as making espresso drinks, folding diverse laundry, and assembling boxes.
• MEM (Multi-Scale Embodied Memory) A memory architecture combining an efficient video-encoder short-term memory (interleaved spatial/temporal attention) with language-based long-term memory, supporting tasks up to 15 minutes long, in-context adaptation, and partial-observability handling.
• RLT (Reinforcement Learning on Tokens / RL Tokens) An encoder-decoder bottleneck that produces a compact RL token from a frozen VLA, enabling on-robot online RL fine-tuning in minutes or hours for precise tasks such as screwdriver alignment and Ethernet insertion without retraining the full model.
• π0.7 (pi0.7) A steerable VLA foundation model that exhibits a step-change in generalization, with emergent capabilities and first signs of compositional generalization enabling the model to follow new language commands and solve tasks not seen in training data, including cross-embodiment transfer such as zero-shot laundry folding on a UR5e bimanual system.

Quantifiable outcome

• Recap RL more than doubled throughput and cut failure rates by 2x or more on espresso-making, box-assembly, and diverse laundry tasks
• +7 more outcomes

Named customers6 records

Segments5 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration5 records

AI capability13 records

Feature10 records

Partnerships

Nine partnerships are on record, tiered minor, flagship and core.

• ProntominorStrategic or Co-development PartnerIndian home-services startup; PI co-founder Lachy Groom personally invested $20M in Pronto's $45M Series B extension at a $200M valuation. Pronto has filmed inside customers' homes to gather data for training physical AI.
• Weave RoboticsflagshipStrategic or Co-development PartnerStrategic deployment partner; Weave integrates PI's VLA models into its home/laundromat folding robots. Joint results show π0.6 reducing missed grasps by 42% and remote interventions by 50% vs π0.5.
• UltraflagshipStrategic or Co-development PartnerStrategic deployment partner; Ultra runs PI's VLA models on its industrial e-commerce packaging robots, reaching 96.4% autonomy for a full shift and improved throughput from π0.5 to π0.6.
• TeleexistencecoreStrategic or Co-development PartnerJapanese robotics company; partner relationship yielded zero-shot motion generation handling unexpected situations in convenience stores, with 30,000+ hours of operational data and a humanoid robot partnership with Seven-Eleven Japan for H1 2026 deployment.
• Robot.com (formerly Kiwibot)flagshipStrategic or Co-development PartnerPartnering with PI to develop custom foundation models for the R-noid wheeled humanoid; R-noid runs on PI's π0.7 VLA model. Commercial deployments include a golf course in New York and food manufacturing customers, with fewer than 40 units deployed across roughly a dozen sites.
• AgiBotcoreStrategic or Co-development PartnerStrategic partnership to pioneer global innovation in embodied intelligence; AgiBot is one of China's leading humanoid robotics companies.
• Scale AIcoreTechnology or IntegrationScale AI's Data Engine for Physical AI serves as a data-collection and annotation partner for PI, leveraging Scale's 100,000+ production hours at its San Francisco prototyping lab and experience with autonomous-vehicle data.
• NVIDIAcoreTechnology or IntegrationPI's models are deployed on NVIDIA robotics hardware/ecosystem (e.g., Isaac, Jetson Thor for VLA inference); PI's work is highlighted at NVIDIA GTC.
• Argonne National Laboratory (RoSA project)coreStrategic or Co-development PartnerGovernment / scientific research partner using PI's VLA models on off-the-shelf humanoid robots in the RoSA (Robot Scientific Assistant) project, targeting 5x improvement in lab automation within a year.

Scale indicators12 records

Recent moves6 records

Expansion highlights8 records

Direct peers

• Skild AI: Direct competitor building a general-purpose robot foundation model (Skild Brain). Frequently cited alongside Physical Intelligence as one of the top-tier physical-AI startups, with comparable funding scale and overlapping target market of robotics deployment partners.
• Figure AI: Direct competitor building AI for humanoid robots (Figure 01/02), backed by Bezos, OpenAI, Microsoft, and NVIDIA at multi-billion-dollar valuations. Competes with PI for both capital and downstream humanoid OEM partnerships.
• Sanctuary AI: Canadian humanoid robotics company building general-purpose AI for dexterous manipulation (Phoenix robot). Targets overlapping industrial and commercial service-robot deployments with a foundation-model approach similar to PI's π family.
• Apptronik: Austin-based humanoid robot company (Apollo) partnering with Google DeepMind on AI and with Mercedes-Benz on industrial pilots. Competes for the same humanoid OEM/intelligence-layer positioning as PI.
• 1X Technologies: Norwegian humanoid robotics company (backed by OpenAI) building its own in-house foundation models for the Neo humanoid. Targets similar consumer/home and industrial use cases as PI's partner deployments.

Regional players

• AgiBot: Leading Chinese humanoid robotics company with which PI has a strategic embodied-intelligence partnership. AgiBot is both a partner and a regional competitor building its own foundation models for the China market.

Broad incumbents

• Boston Dynamics: Established humanoid/quadruped robot manufacturer (Atlas, Spot) now owned by Hyundai. Historically focused on model-predictive control but moving toward learned policies, making it a broad incumbent competitor in dynamic robot autonomy.
• Covariant: Founded by Pieter Abbeel (Berkeley, like PI co-founder Sergey Levine) to build foundation models for warehouse picking/packing robots; acquired by Amazon in 2024. Competes head-on with PI's Ultra-style warehouse manipulation use case but is now inside Amazon.
• Google DeepMind Robotics: DeepMind (RT-2, RT-X, Gemini Robotics) is the leading internal competitor building VLA models for robot control. PI's founders came from DeepMind, and DeepMind benefits from Alphabet/Google compute and data — making it the most credible broad incumbent PI must outpace.
• NVIDIA (GR00T / Isaac): NVIDIA is both a technology partner (Jetson Thor, Isaac, GTC showcase) and a competitor via GR00T foundation models for humanoid robots. Owns the dominant compute and simulator stack that PI's models run on, giving it vertical leverage.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat7 records

Key risks7 records

Key highlights7 records

Customer concentration

• Xx.com/physical_int
• GitHubgithub.com/Physical-Intelligence/openpi
• LinkedInlinkedin.com/company/physical-intelligence

Revenue estimate

Valuation estimate

Number of profiles

Profiles7 records

Funding overview

Funding rounds4 records

Investors26 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Physical Intelligence