> ## Content Index
> Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Physical Intelligence: A Full Guide to pi0, pi0.5, the Robot Foundation Model Bet and Its Limits
- URL: https://www.insidedeeptech.com/physical-intelligence-pi0-robot-foundation-model-full-guide/
- Published: 2026-11-20T07:13:32.000Z
- Updated: 2026-11-20T07:13:31.000Z
- Description: Physical Intelligence builds the pi family of robot foundation models. This guide covers pi0, pi0.5, pi*0.6 and pi0.7, openpi, funding up to the reported $20 billion talks, the Pi API beta, rivals and honest limits.
- Author: Austin Heaton
- Tags: Robotics, Physical AI, AI, Deep Tech

Physical Intelligence doesn't sell a robot. It doesn't sell a product you can buy off a price list either. Yet in October 2026 Bloomberg reported the San Francisco lab was in talks to raise [at least $1 billion at a valuation of roughly $20 billion](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com), led by existing backers Avenir and Lux Capital.

The bet behind that number is simple to say and very hard to prove: one family of robot foundation models, π0 (pi0), π0.5, π\*0.6 and π0.7, that can learn to control almost any robot body for almost any task.

This guide covers each model release, how vision-language-action (VLA) flow-matching models actually work, the openpi open-source code, the funding trail, the new Pi API, how Physical Intelligence compares with Google DeepMind, Skild AI, NVIDIA, Figure and Tesla, and the honest limits of what's been shown. Last updated: November 20, 2026.

## Key Takeaways

- Physical Intelligence was reported in October 2026 to be raising [at least $1 billion at about $20 billion](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com).
- Its first model, π0, pairs a 3 billion parameter VLM with a [300M action expert for 3.3 billion parameters](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com).
- π0, π0-FAST and π0.5 weights are open source in [openpi under Apache-2.0](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com).
- π\*0.6 used RL to [more than double throughput on its hardest tasks](https://arxiv.org/abs/2511.14759?ref=insidedeeptech.com).
- The [Pi API beta](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com) opened October 13, 2026, with no public pricing or production SLA.

## Physical Intelligence at a glance

Here's the short version. Capability claims are the company's own lab results.

| Item                 | Detail                                                                                                |
| -------------------- | ----------------------------------------------------------------------------------------------------- |
| Founded              | 2024, San Francisco                                                                                   |
| Co-founders          | Karol Hausman (CEO), Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail, Quan Vuong |
| What it builds       | Robot foundation models (VLAs), not robot hardware                                                    |
| Model family         | π0, π0-FAST, π0.5, π\*0.6, π0.7                                                                       |
| Open source          | openpi repo: π0, π0-FAST, π0.5 weights and code                                                       |
| Commercial access    | Pi API beta (Oct 13, 2026), application based                                                         |
| Last confirmed round | Series B, $600 million at $5.6 billion (Nov 2025, reported)                                           |
| Filed round          | Series C, about $1.05 billion at $11 billion to $11.2 billion (May 2026 filings)                      |
| Talks reported       | At least $1 billion at about $20 billion (Oct 2026)                                                   |

Source: [π0 blog](https://www.pi.website/blog/pi0?ref=insidedeeptech.com) and [openpi](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com) for the models; [SiliconANGLE](https://siliconangle.com/2025/11/20/jeff-bezos-backed-physical-intelligence-raises-600m-improve-ai-robot-brains/?ref=insidedeeptech.com) for the $600 million Series B; [Revenue Memo](https://www.revenuememo.com/p/physical-intelligence-funding?ref=insidedeeptech.com) for the May 2026 Series C filings; [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) for the $20 billion talks; [Pi API announcement](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com) for the beta.

💡

**What Physical Intelligence sells, in one line:** a general-purpose robot brain. Its models take camera images, a language instruction and the robot's joint state, and output motor commands. Partners bring the hardware and the task data. As of November 20, 2026, the company has no public price list.

## Why Physical Intelligence exists: robots are narrow specialists

The company's founding argument is Moravec's paradox. Chess and drug discovery are easy for AI. Folding a shirt is hard. Today's industrial robots repeat one weld in one spot, and every new behavior needs its own engineering and its own data.

The π0 launch post puts the data problem bluntly: language models mine the web, but [there's no equivalent treasure trove of robot data](https://www.pi.website/blog/pi0?ref=insidedeeptech.com). So the plan is a single generalist policy, trained on many robots and tasks, that only needs a little extra data for each new application.

Co-founder Sergey Levine described it to TechCrunch as [ChatGPT, but for robots](https://techcrunch.com/2026/03/27/physical-intelligence-is-reportedly-in-talks-to-raise-1-billion-again/?ref=insidedeeptech.com). Co-founder Lachy Groom told the same outlet there's no timeline for commercialization.

If you want the wider category context first, Inside Deep Tech's [robot foundation models explainer](https://www.insidedeeptech.com/robot-foundation-models-the-ai-brains-behind-physical-ai/) covers how these models fit into physical AI as a whole.

## The π model timeline: from π0 to π0.7

Physical Intelligence publishes research on a steady cadence. Here's every major model and method release on its [research blog](https://www.pi.website/blog?ref=insidedeeptech.com) through November 2026.

| Date         | Release              | What changed                                                    |
| ------------ | -------------------- | --------------------------------------------------------------- |
| Oct 31, 2024 | π0                   | First generalist policy; flow-matching action expert on a VLM   |
| Jan 16, 2025 | FAST tokenizer       | Compresses action sequences so autoregressive VLAs train faster |
| Feb 4, 2025  | openpi (π0, π0-FAST) | Open weights and fine-tuning code                               |
| Apr 22, 2025 | π0.5                 | Open-world generalization to unseen homes                       |
| Nov 17, 2025 | π\*0.6 with Recap    | RL from autonomous experience and corrections                   |
| Apr 16, 2026 | π0.7                 | Steerable model with compositional generalization               |
| Oct 13, 2026 | Pi API beta          | Hosted fine-tuning and inference for partners                   |

Source: [Physical Intelligence blog index](https://www.pi.website/blog?ref=insidedeeptech.com), checked November 20, 2026.

### π0: the first generalist policy

π0 starts from PaliGemma, an open vision-language model. The team adds a separate "action expert" and gets a [3 billion parameter VLM plus 300M action parameters, 3.3 billion in total](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com). The VLM brings web-scale knowledge of objects and language. The action expert turns that into motion.

The pre-training mixture spans [7 robot configurations and 68 tasks](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com), with open datasets such as Open X-Embodiment, Bridge and DROID making up 9.1% of the mix by timestep. Robots in the set range from single UR5e and Franka arms to bimanual ALOHA-style rigs and mobile manipulators.

The headline demos were laundry folding from a dryer, table bussing and box assembly. Post-training for those tasks used anywhere from [5 hours to 100 or more hours](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com) of task data, depending on difficulty.

### π0-FAST: actions as tokens

FAST (Frequency-space Action Sequence Tokenization) compresses action chunks with a discrete cosine transform, so a VLA can predict actions as text-like tokens. The paper says FAST with π0 can train on 10k hours of robot data and [match diffusion VLAs while cutting training time by up to 5x](https://arxiv.org/abs/2501.09747?ref=insidedeeptech.com). It also released FAST+, a tokenizer trained on 1M real robot action trajectories.

The catch is inference speed. In the openpi release notes, Physical Intelligence says π0-FAST follows language a bit better but has [about 4 to 5x higher inference cost](https://www.pi.website/blog/openpi?ref=insidedeeptech.com) than flow-matching π0.

### π0.5: generalizing to homes it has never seen

π0.5 kept the π0 architecture and changed the recipe. It co-trains on robot demonstrations from many robot types, high-level subtask labels, verbal instructions and web data such as captioning and object detection. The [π0.5 paper](https://arxiv.org/abs/2504.16054?ref=insidedeeptech.com) says this is the first time an end-to-end learned system cleaned a kitchen or bedroom in entirely new homes.

The company's own ablations matter here. In the [π0.5 post](https://www.pi.website/blog/pi05?ref=insidedeeptech.com), removing web data hurt performance most on unfamiliar objects, while removing data from other robots hurt across every test condition. Physical Intelligence also says plainly that π0.5 "is far from perfect" and makes both reasoning and motor mistakes.

### π\*0.6 and Recap: learning from experience

Imitation learning gets a robot to work some of the time. It rarely gets it to work all day. Recap (RL with Experience and Corrections via Advantage-conditioned Policies) adds two more data sources to demonstrations: expert corrections during autonomous runs and the robot's own experience.

On the hardest tasks, the [π\*0.6 paper](https://arxiv.org/abs/2511.14759?ref=insidedeeptech.com) says Recap more than doubles throughput and roughly halves the failure rate. The company's [π\*0.6 post](https://www.pi.website/blog/pistar06?ref=insidedeeptech.com) shows the model making espresso drinks from 5:30am to 11:30pm, folding 50 novel laundry items in a new home and assembling 59 chocolate boxes in a real factory.

### π0.7: one steerable model instead of many specialists

π0.7, published April 16, 2026, is the most ambitious claim so far. It's trained with richer prompts: language, metadata about how well or how fast to do a task, control modes and visual subgoal images. That lets the model absorb messy data, including suboptimal autonomous runs, without learning bad habits.

The standout example is cross-embodiment transfer. With no laundry data on a bimanual UR5e system, π0.7 folded shirts on it at a success rate that matched expert teleoperators trying that robot for the first time, operators with a [mean of 375 hours of teleoperation experience](https://www.pi.website/blog/pi07?ref=insidedeeptech.com). The same post says one π0.7 model matched the RL-trained π\*0.6 specialists on their own tasks.

⚠️

**Read the π0.7 claims carefully.** Every number comes from Physical Intelligence's own evaluations on its own robots and tasks. There is no shared public benchmark where π0.7, Gemini Robotics and GR00T are scored side by side.

## How a VLA flow-matching model actually works

Let's start with the inputs. Each step, π0 sees one or more camera images, the language command and the robot's proprioceptive state (joint angles, gripper position). Images and text go through the VLM backbone; state and actions go through the smaller action expert. Both share one transformer and talk through attention.

Here's the problem a plain VLM has: it outputs discrete text tokens, but a robot arm needs smooth continuous commands many times per second. π0 solves this with flow matching, a close cousin of diffusion models, so it can output [motor commands up to 50 times per second](https://www.pi.website/blog/pi0?ref=insidedeeptech.com).

- **Action chunks:** the model predicts a chunk of [H=50 future actions](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com) at once, so the robot can keep moving between inference calls.
- **Denoising:** it starts from random noise and refines the chunk over [10 integration steps](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com), reusing cached attention for the image and text prefix.
- **Two kinds of training:** broad pre-training gives general skills; short post-training on high-quality data specializes the model for a task.
- **FAST alternative:** π0-FAST swaps flow matching for compressed action tokens, which trains faster but runs slower.

## openpi: what's actually open source

Physical Intelligence released [π0 weights and code in February 2025](https://www.pi.website/blog/openpi?ref=insidedeeptech.com), then added π0.5 and PyTorch support in September 2025\. The [openpi GitHub repo](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com) is Apache-2.0 licensed and had more than 14,700 stars as of November 20, 2026.

The repo ships base checkpoints pre-trained on [10k+ hours of robot data](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com), plus fine-tuned demos for ALOHA, DROID and the Libero benchmark. The company says [1 to 20 hours of data](https://www.pi.website/blog/openpi?ref=insidedeeptech.com) was enough to fine-tune to a range of tasks in its own tests, with a clear warning that results on your robot may vary.

| openpi mode        | GPU memory needed | Example GPU       |
| ------------------ | ----------------- | ----------------- |
| Inference          | More than 8 GB    | RTX 4090          |
| Fine-tuning (LoRA) | More than 22.5 GB | RTX 4090          |
| Fine-tuning (full) | More than 70 GB   | A100 80GB or H100 |

Source: [openpi README](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com), checked November 20, 2026\. Ubuntu 22.04 is the only tested OS, and π0.7 is not in the open repo.

Here's why that matters for you. openpi gives any lab a credible VLA baseline without training from scratch. But the newest models, π\*0.6 and π0.7, stay closed. The open tier is a research on-ramp, and the commercial tier is the API.

## The Pi API beta and early partners

On October 13, 2026, Physical Intelligence opened the [Pi API beta](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com). The workflow is three steps: upload your robot data, launch fine-tuning on compute Physical Intelligence manages, and call the fine-tuned model through a hosted inference gateway.

Two partners get detailed write-ups. Actor in Utah runs an autonomous wheel loader that loads dump trucks with no operator in the cab. Almond in San Francisco says its Axol dual-arm robot deburred its own spacers after iterating with the API for four days. Most showcased deployments, from Japan to Switzerland to the US, used cloud inference, and the company admits [network connectivity was a challenge for some](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com).

Access is application based, with teams onboarded in batches, according to [Humanoids Daily](https://www.humanoidsdaily.com/news/physical-intelligence-api-beta-robot-training-inference?ref=insidedeeptech.com). No pricing, uptime commitment, hardware compatibility list or local inference option has been published.

## Funding, valuation and investors

The valuation curve is one of the steepest in robotics. Physical Intelligence raised a [$70 million seed in early 2024](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com), then a [$400 million round at a $2.4 billion valuation](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com) in November 2024 led by Jeff Bezos, Thrive Capital and Lux Capital. OpenAI, Sequoia and Khosla Ventures were early backers.

| Round                         | Date       | Amount                                                                                                                                                                  | Valuation                                                                                                                                                             | Lead or notable investors                         |
| ----------------------------- | ---------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------- |
| Seed                          | Early 2024 | [$70 million](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com)                              | Not disclosed                                                                                                                                                         | Thrive, Khosla, Lux, OpenAI, Sequoia              |
| Series A                      | Nov 2024   | [$400 million](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com)                             | [$2.4 billion](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com)                           | Jeff Bezos, Thrive, Lux                           |
| Series B                      | Nov 2025   | [$600 million](https://siliconangle.com/2025/11/20/jeff-bezos-backed-physical-intelligence-raises-600m-improve-ai-robot-brains/?ref=insidedeeptech.com)                 | [$5.6 billion](https://siliconangle.com/2025/11/20/jeff-bezos-backed-physical-intelligence-raises-600m-improve-ai-robot-brains/?ref=insidedeeptech.com)               | CapitalG lead; Index, T. Rowe Price, Bezos        |
| Series C (filed, unannounced) | May 2026   | [About $1.05 billion](https://www.revenuememo.com/p/physical-intelligence-funding?ref=insidedeeptech.com)                                                               | [$11 billion to $11.2 billion](https://www.revenuememo.com/p/physical-intelligence-funding?ref=insidedeeptech.com)                                                    | Founders Fund, Lightspeed, Thrive, Lux (reported) |
| Talks reported                | Oct 2026   | [At least $1 billion](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) | [About $20 billion](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) | Avenir, Lux (in talks to lead)                    |

Source: [The Robot Report](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com) for seed and Series A; [SiliconANGLE](https://siliconangle.com/2025/11/20/jeff-bezos-backed-physical-intelligence-raises-600m-improve-ai-robot-brains/?ref=insidedeeptech.com) citing Bloomberg and The Information for Series B; [TechCrunch](https://techcrunch.com/2026/03/27/physical-intelligence-is-reportedly-in-talks-to-raise-1-billion-again/?ref=insidedeeptech.com) and [Revenue Memo](https://www.revenuememo.com/p/physical-intelligence-funding?ref=insidedeeptech.com) for the Series C talks and filings; [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) for the October 2026 talks.

Two details matter. First, Physical Intelligence hasn't officially announced its Series B or Series C; both come from press reports and corporate filings. Second, Alphabet's [CapitalG](https://capitalg.com/insights/physical-intelligence-bringing-general-purpose-ai-into-the-physical-world/?ref=insidedeeptech.com) led the Series B even though Google DeepMind builds a direct competitor, Gemini Robotics.

TechCrunch reported in March 2026 that the company had [about 80 employees](https://techcrunch.com/2026/03/27/physical-intelligence-is-reportedly-in-talks-to-raise-1-billion-again/?ref=insidedeeptech.com) at the start of the year. That's a tiny team for a company priced like a frontier lab, and most of the money goes to compute and robot data collection.

## Physical Intelligence vs Gemini Robotics, Skild AI, GR00T, Helix and Tesla

| Company               | Model               | Openness                    | Hardware strategy          | Latest public evidence                                                                                                                                              |
| --------------------- | ------------------- | --------------------------- | -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Physical Intelligence | π0 to π0.7          | π0/π0.5 open; newest closed | Model only, partner robots | π0.7 cross-embodiment results (Apr 2026)                                                                                                                            |
| Google DeepMind       | Gemini Robotics 2.5 | Closed, partner rollout     | Partners incl. Apptronik   | [96% success in Mercedes-Benz kitting pilot](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/?ref=insidedeeptech.com) |
| Skild AI              | Skild Brain         | Closed                      | Model plus deployments     | [$1.4 billion Series C at over $14 billion](https://www.skild.ai/blogs/series-c?ref=insidedeeptech.com)                                                             |
| NVIDIA                | Isaac GR00T N1.7    | Open, commercial license    | Platform for any humanoid  | [3B params, 20,854 hours of human video](https://huggingface.co/blog/nvidia/gr00t-n1-7?ref=insidedeeptech.com)                                                      |
| Figure AI             | Helix               | Closed                      | Own humanoid (Figure 03)   | [7B System 2 plus 80M System 1](https://www.figure.ai/news/helix?ref=insidedeeptech.com)                                                                            |
| Tesla                 | Optimus stack       | Closed                      | Own humanoid (Optimus)     | Few public model details                                                                                                                                            |

Source: [Google DeepMind](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/?ref=insidedeeptech.com), [Skild AI](https://www.skild.ai/blogs/series-c?ref=insidedeeptech.com), [NVIDIA on Hugging Face](https://huggingface.co/blog/nvidia/gr00t-n1-7?ref=insidedeeptech.com), [Figure](https://www.figure.ai/news/helix?ref=insidedeeptech.com); Physical Intelligence data from its [π0.7 post](https://www.pi.website/blog/pi07?ref=insidedeeptech.com).

**Google DeepMind** is the closest research rival. Gemini Robotics 2.5, announced November 10, 2026, was tested on [over 200 tasks with 30% to 100% higher success rates](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/?ref=insidedeeptech.com) than earlier baselines. Its Apptronik deployment work is covered in Inside Deep Tech's [Apptronik Apollo 2 guide](https://www.insidedeeptech.com/apptronik-apollo-2-modular-humanoid-robot-park-full-guide/).

**Skild AI** is the closest business rival. It also pitches one brain for any robot, but it leans on simulation, internet video and its own deployments, and says live revenue went from [zero to about $30M in a few months of 2025](https://www.skild.ai/blogs/series-c?ref=insidedeeptech.com). Physical Intelligence hasn't disclosed any revenue.

**NVIDIA** plays a different game. GR00T N1.7 is open and commercially licensed, so it sells GPUs and Jetson Thor modules whether or not its own model wins. Inside Deep Tech's [Isaac GR00T guide](https://www.insidedeeptech.com/understanding-nvidia-isaac-gr00t-a-foundation-model-for-generalist-robots/) covers the platform.

**Figure and Tesla** build vertically: their own humanoid, their own model, their own factories. Figure's Helix runs fully onboard low-power GPUs, per [Figure](https://www.figure.ai/news/helix?ref=insidedeeptech.com). For how those two compare on hardware, see the [Figure 03 vs Tesla Optimus comparison](https://www.insidedeeptech.com/figure-03-vs-tesla-optimus-2026-humanoid-robot-comparison/). Physical Intelligence is betting that a hardware-agnostic model wins the same way Android beat dozens of bespoke phone operating systems.

## Honest limits: what Physical Intelligence hasn't shown yet

- **No independent benchmark.** Every π0.5, π\*0.6 and π0.7 result comes from the company's own tasks and scoring.
- **No disclosed revenue or customers at scale.** The API is a beta with no published price, against Skild's reported [revenue run-up](https://www.skild.ai/blogs/series-c?ref=insidedeeptech.com).
- **Cloud latency.** Most API showcase deployments used cloud inference, and the company flags [network connectivity issues](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com).
- **Data is still mostly teleoperation.** The [π\*0.6 post](https://www.pi.website/blog/pistar06?ref=insidedeeptech.com) says today's models rely on human-collected demonstrations, which cap speed and cost.
- **Safety is thin in public.** DeepMind published a robotics jailbreak benchmark with Gemini Robotics 2.5; Physical Intelligence's research blog lists no comparable safety benchmark release.
- **Valuation runs far ahead of evidence.** A reported [$20 billion price](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) for a company with about 80 staff and no product price list leaves little room for slips.

🧭

**Who this doesn't work for yet:** a buyer that needs a guaranteed cycle time, an SLA and an on-premise model next quarter. Today, Physical Intelligence is a research partner and a beta API, not a turnkey automation vendor.

## Inside Deep Tech's take

Inside Deep Tech's take: Physical Intelligence has the strongest published research record of any robot foundation model startup. π0 set the architecture template, openpi made it the default open baseline, and π0.7's cross-robot transfer is the clearest evidence yet that robot data can compound the way web text did.

But a model company has to win distribution. Google has Gemini and Apptronik, NVIDIA has the chips, and Figure and Tesla own their robots. The Pi API is the first real test of whether robot makers will rent a brain rather than build one.

Inside Deep Tech's take: expect the next 12 months to be judged on API customers and uptime, not demo videos.

## FAQ

#### What is Physical Intelligence?

Physical Intelligence is a San Francisco AI company founded in 2024 that builds general-purpose robot foundation models, the π family, rather than robot hardware. Its first model was [π0 in October 2024](https://www.pi.website/blog/pi0?ref=insidedeeptech.com).

#### What is π0 (pi0)?

π0 is a vision-language-action model that adds a flow-matching action expert to the PaliGemma VLM, for [3.3 billion parameters in total](https://arxiv.org/abs/2410.24164?ref=insidedeeptech.com). It outputs continuous motor commands for many robot types.

#### Is Physical Intelligence open source?

Partly. π0, π0-FAST and π0.5 weights and code are in the [openpi repo under Apache-2.0](https://github.com/Physical-Intelligence/openpi?ref=insidedeeptech.com). Newer models such as π\*0.6 and π0.7 are not open.

#### How much is Physical Intelligence worth?

Its last reported priced round valued it at [$11 billion to $11.2 billion](https://www.revenuememo.com/p/physical-intelligence-funding?ref=insidedeeptech.com) in May 2026\. Bloomberg reported in October 2026 that it was in talks at [about $20 billion](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com).

#### Who invests in Physical Intelligence?

Reported backers include Jeff Bezos, Thrive Capital, Lux Capital, OpenAI, Sequoia, Khosla Ventures, CapitalG, Index Ventures, T. Rowe Price, Founders Fund and Lightspeed, per [The Robot Report](https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/?ref=insidedeeptech.com), [SiliconANGLE](https://siliconangle.com/2025/11/20/jeff-bezos-backed-physical-intelligence-raises-600m-improve-ai-robot-brains/?ref=insidedeeptech.com) and [TechCrunch](https://techcrunch.com/2026/03/27/physical-intelligence-is-reportedly-in-talks-to-raise-1-billion-again/?ref=insidedeeptech.com).

#### Can I use Physical Intelligence models on my robot?

Yes, two ways. You can fine-tune openpi checkpoints yourself on an NVIDIA GPU, or apply to the [Pi API beta](https://www.pi.website/blog/api-beta?ref=insidedeeptech.com) for hosted fine-tuning and inference.

## What to watch next

- Whether the reported [$20 billion round](https://www.bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation?ref=insidedeeptech.com) closes, and who leads it.
- Named Pi API customers, pricing and any on-robot (local) inference option.
- A π0.7 or successor release in openpi, or a shift away from open weights.
- Third-party benchmarks that put π models next to Gemini Robotics and GR00T.
- How Physical Intelligence ranks in the next update of Inside Deep Tech's [Top 100 Physical AI Companies](https://www.insidedeeptech.com/top-100-physical-ai-companies/).

---

If you're evaluating a robot brain for a 2027 pilot, start with openpi on a test cell, then compare it against GR00T and any Gemini Robotics partner access before you sign an API deal. The [Top 100 Physical AI Companies](https://www.insidedeeptech.com/top-100-physical-ai-companies/) directory is the fastest way to map the rest of the field.

[Browse the Top 100 Physical AI Companies](https://www.insidedeeptech.com/top-100-physical-ai-companies/)

[Robot Foundation Models: The AI Brains Behind Physical AIHow robot foundation models are trained, what they can do and where they still break.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/robot-foundation-models-the-ai-brains-behind-physical-ai/)

Cover image: generated robot arms in a lab, blurred with title overlay for this post.