Physical Intelligence (π), the robotics company backed by Bond, Khosla Ventures, Sequoia Capital, OpenAI, and Jeff Bezos, released the Pi API beta on October 13, 2026. The API lets developers upload their own demonstration data, fine-tune Physical Intelligence’s foundation models on managed compute infrastructure, and deploy the resulting policies on their own robots through a hosted inference gateway. For robotics builders who have been watching foundation models reshape language and image generation, this is the first commercially available API that extends that paradigm to physical robot control — and it raises real questions about who benefits, what the limitations are, and how quickly the robotics ecosystem will adapt.

In Brief

  • Physical Intelligence launched the Pi API beta on October 13, 2026, offering fine-tuning and cloud-hosted inference for its robotics foundation models
  • The company’s model lineage runs from π0 (October 2024, open-sourced February 2025) through π0.5 (April 2025) to π0.7 (April 2026), with each version improving generalization and steerability
  • Early API partners span construction (Actor’s autonomous wheel loaders), agriculture (AgRobotics), food service (Chef Robotics, CloudChef), and manufacturing (Almond’s dual-arm Axol robot)
  • The open-source models are available under Apache 2.0 on GitHub, requiring an NVIDIA GPU with at least 8 GB VRAM for inference and 22.5 GB for LoRA fine-tuning
  • Cloud-hosted inference means robots can run policies without local GPU hardware, though network connectivity becomes a dependency
  • The API is in beta, with access gated by a waitlist and application process

From π0 to the Pi API: A Quick Timeline

Physical Intelligence introduced π0, its first generalist robot policy, on October 31, 2024. The model, described in a paper published at RSS 2025, used a flow matching architecture built on top of a pre-trained vision-language model to inherit internet-scale semantic knowledge. It was trained on data from multiple dexterous robot platforms — single-arm, dual-arm, and mobile manipulators — and could perform tasks like laundry folding, table cleaning, and box assembly in zero-shot conditions after pre-training.

The company open-sourced π0 on February 4, 2025, releasing model weights and code under an Apache 2.0 license through the Physical-Intelligence/openpi GitHub repository. This was a significant move: it gave academic researchers and independent builders access to a production-grade robotics foundation model without requiring partnership agreements or proprietary licenses. The repository also included π0-FAST, an autoregressive variant based on a new action tokenizer that reportedly trained five times faster than the original model.

In April 2025, π0.5 extended the architecture with open-world generalization. According to the paper on arXiv, π0.5 used co-training on heterogeneous tasks — combining data from multiple robots, web data, and hybrid multi-modal examples — to enable manipulation in entirely new environments. The company demonstrated the model cleaning kitchens and bedrooms in homes it had never encountered.

By April 2026, π0.7 arrived, described by the company as a “steerable model with emergent capabilities.” The Pi website characterizes it as a step-change in generalization, though the accompanying paper is not yet available on arXiv for independent verification.

The API beta, announced on October 13, 2026, is the logical next step: turning a research model series into a commercial platform.

What the Pi API Actually Does

The API workflow has three stages, according to the company’s announcement post.

Upload. Developers collect demonstration data — typically teleoperated recordings of a robot performing a task — and upload it through the API. This data becomes the basis for fine-tuning a foundation model to the developer’s specific robot, environment, and task.

Fine-tune. Physical Intelligence manages the compute infrastructure for fine-tuning. The developer does not need to provision GPUs, configure training pipelines, or handle the machine learning infrastructure that would normally be required. The company states that the latest model available through the API “already provides more data-efficient fine-tuning and improved capability” compared to earlier versions.

Deploy. Fine-tuned models are accessed through a hosted inference gateway. The robot sends observations (camera images, joint states) to the cloud, and the model returns actions. Most early deployments shown in the company’s announcement ran with cloud-hosted inference, though the company acknowledged “network connectivity challenges for some.”

This is structurally similar to how OpenAI’s API made GPT models accessible to developers: you bring your data, the platform handles the ML infrastructure, and you consume the model through an API. The difference is that the input and output are physical — camera frames and robot commands — rather than text.

What the Early Deployments Reveal

Physical Intelligence shared video demonstrations of partners using the API across several industries. These are company-selected examples, not independent evaluations, but they illustrate the range of tasks the models are being adapted to.

Construction. Actor, a company deploying autonomous heavy machinery, used the API to fine-tune a model for a wheel loader that scoops material and loads it into a dump truck during what the company describes as a real commercial deployment — without an operator inside the machine.

Food service. Chef Robotics and CloudChef are using the API for food assembly tasks. Almond, a startup building Axol — a dual-arm robot for “physical AI builders” — used the API to fine-tune a model in four days that deburrs metal spacers used in the robot’s own assembly. This is a notable example because it shows the API enabling a robot to help manufacture itself.

Agriculture. AgRobotics is deploying models for agricultural tasks, though the specific operations were not detailed.

Manufacturing. Deft, Formic Robotics, Foundry, and Frontier Machines are listed as partners, suggesting applications in warehouse automation and industrial manipulation.

The diversity matters because robotics foundation models have historically struggled with task diversity. A model that can fold laundry, load dump trucks, and deburr metal parts — even after task-specific fine-tuning — represents a meaningful expansion of what a single model family can do.

Open-Source Models vs. the Commercial API

One of the most interesting aspects of Physical Intelligence’s approach is that the company is offering both an open-source path and a commercial API.

The openpi repository on GitHub contains π0, π0-FAST, and π0.5 model checkpoints under an Apache 2.0 license. Developers with an NVIDIA GPU — at least 8 GB VRAM for inference, 22.5 GB for LoRA fine-tuning, or 70 GB for full fine-tuning — can run and adapt these models locally without paying for API access. The repository includes training configs, fine-tuning examples, and support for PyTorch.

The commercial API, by contrast, handles the infrastructure burden. For teams that do not have A100 or H100 GPUs, do not want to manage training pipelines, or need the latest model versions (which may not be immediately open-sourced), the API is the practical path. It also provides cloud-hosted inference, which means the robot itself does not need an onboard GPU — a significant consideration for cost-sensitive deployments.

This dual approach creates a natural segmentation: researchers and well-resourced teams use the open-source models; commercial deployments and teams without ML infrastructure use the API. Whether the latest models will always be open-sourced, or whether the API will eventually offer capabilities that the open-source models do not, remains an open question.

What This Means for Product Teams

For product teams building robots, the Pi API changes the economics of robot intelligence in three ways.

Reduced data requirements. Foundation models pre-trained on 10,000+ hours of robot data — the figure cited in the openpi documentation — provide a starting point that substantially reduces the amount of task-specific data needed. The company states that its latest API model “already provides more data-efficient fine-tuning,” though specific data requirements are not published. For comparison, training a manipulation policy from scratch typically requires hundreds to thousands of demonstrations per task.

Eliminated ML infrastructure burden. Fine-tuning a vision-language-action model requires significant GPU resources and ML engineering expertise. The API abstracts this away. A team with robot hardware and demonstration data can obtain a working policy without hiring an ML engineer or provisioning training infrastructure.

Faster iteration cycles. The API enables a collect-train-test loop where developers gather demonstrations, launch fine-tuning, test on the robot, and use failures to guide the next data collection round. Almond’s four-day turnaround from data collection to a working deburring policy illustrates the speed this enables.

The trade-off is dependency on Physical Intelligence’s infrastructure. Cloud-hosted inference introduces latency and requires network connectivity — a challenge the company itself acknowledged. For applications in remote areas, underground environments, or settings with unreliable internet, local inference using the open-source models may be necessary despite the higher infrastructure requirements.

What This Means for Educators and Emerging Markets

For STEAM educators and robotics programs, the Pi API and the open-source models create new teaching opportunities — and new challenges.

The open-source models provide a concrete example of how modern AI techniques (transformers, vision-language models, flow matching) are being applied to physical systems. Students who have learned about large language models can see how similar architectures extend to robot control. The openpi repository includes fine-tuning examples that could serve as advanced course material for university-level robotics or AI programs.

For platforms like RoboSim, which lets students program virtual robots in a 3D environment, and LearnOSTEAM, which combines interactive lessons with a coding environment, foundation model APIs raise a design question: should educational platforms incorporate real foundation model inference so students can experiment with the same tools used in production robotics? The cost of API calls and the need for cloud connectivity make this challenging in practice, particularly for schools in Pakistan and other emerging markets where internet infrastructure remains uneven.

The educational robotics platform guide we published recently discusses how simulation environments scale to large class sizes at lower cost than hardware kits. Foundation model APIs could eventually make advanced AI capabilities accessible in simulation — but only if pricing models become education-friendly and offline inference becomes feasible on classroom hardware.

For Pakistani technology teams specifically, the open-source models are the more relevant path. Running inference on a local GPU avoids API latency, sidesteps data sovereignty concerns, and works with the kind of hardware that a university lab or startup in Islamabad might realistically access. The barrier is the 8 GB VRAM minimum for inference — achievable with a used RTX 3060 or similar card, but not with the low-power SBCs that many educational robots use.

Limitations and Open Questions

Cloud inference dependency. The API’s default deployment model requires network connectivity. For industrial robots in factories, this may be acceptable. For mobile robots in variable environments, it creates a single point of failure. The open-source models provide a fallback, but only for teams with adequate GPU hardware.

No published benchmarks for the API. Physical Intelligence has shared partner demonstrations but has not published quantitative benchmarks comparing API-fine-tuned models against baselines on standardized tasks. The π0 paper includes evaluation on tasks like laundry folding and table cleaning, but the API’s fine-tuning performance on novel tasks is documented only through partner testimonials.

Pricing is not public. The API is in beta with a waitlist. No pricing information has been announced. For commercial deployments, the cost of fine-tuning runs, per-inference pricing, and minimum commitments will determine whether the API is viable for cost-sensitive applications.

Robot hardware compatibility. The openpi documentation notes that π0 “was developed for our own robots, which differ from the widely used platforms such as ALOHA and DROID” and that “we do not expect every such attempt to be successful.” The API may handle a broader range of hardware, but compatibility with specific robot platforms remains an empirical question.

Generalization claims need independent verification. The π0.7 model is described as having “emergent capabilities” and a “step-change in generalization,” but without a published paper or independent evaluation, these claims should be treated as company statements rather than confirmed results. The π0 and π0.5 papers, by contrast, are peer-reviewed and available on arXiv.

How This Fits the Broader Robotics Landscape

The Pi API arrives amid a wave of robotics investment and platform development. Boston Dynamics recently opened RMAC at Hyundai’s factory to train Atlas humanoid robots for manufacturing. Figure’s Helix 2.5 demonstrated zero-shot generalization across 30 unseen homes. NVIDIA’s Isaac ROS 5.0 brought AI agent workflows to ROS 2 development. Goldman Sachs has forecast 6.5 million humanoid robots by 2035.

What distinguishes Physical Intelligence’s approach is the combination of open-source foundation models with a commercial fine-tuning API. Boston Dynamics is building robots. Figure is building robots. NVIDIA is building tools for robot developers. Physical Intelligence is building the model layer that could sit underneath all of them — if the models work well enough across diverse hardware.

The closest analogy is not to any robotics company but to OpenAI’s early API strategy: release a powerful general model, let developers adapt it to their specific use cases, and iterate based on what works. The question is whether robotics, with its physical-world constraints and hardware diversity, will follow the same adoption curve as text generation.

What to Watch Next

  • π0.7 paper publication. When the π0.7 paper appears on arXiv, independent researchers can evaluate the generalization claims. Until then, the “step-change in generalization” is a company statement.
  • Pricing announcement. The beta is free for accepted participants, but commercial pricing will determine whether the API is viable for startups and teams in emerging markets.
  • Independent benchmarks. Standardized evaluations on benchmarks like DROID or LIBERO would let the community compare API-fine-tuned models against open-source baselines and competing approaches.
  • Education-tier access. If Physical Intelligence offers discounted or free API access for educational use — similar to how OpenAI and Google provide academic credits — it could make foundation model robotics accessible to university courses and robotics competitions worldwide.
  • Offline inference improvements. The ESP32-S3 BitNet cluster and NVIDIA’s PAIR distributed inference represent alternative approaches to running AI models on constrained hardware. Whether similar compression or distribution techniques can apply to vision-language-action models will determine how far foundation model robotics can reach into low-cost hardware.

Conclusion

Physical Intelligence’s Pi API is a genuine milestone: the first commercial API for fine-tuning and deploying robotics foundation models across diverse hardware. The open-source models provide a research-grade fallback for teams with GPU access, while the API targets commercial deployments that need managed infrastructure. For product teams, educators, and emerging-market builders, the key question is not whether foundation models will matter for robotics — that is increasingly clear — but how quickly the practical barriers of cost, connectivity, and hardware compatibility will come down.

If you are building robots or teaching robotics, the openpi repository is worth exploring today. The API waitlist is worth joining if you have a concrete task and demonstration data ready. And if you are an educator thinking about how to bring foundation model concepts into the classroom, the intersection of simulation platforms like RoboSim and real model APIs is a space worth watching closely.

What task would you want a robot to learn — and do you have the demonstration data to teach it?

Sources

Shamyl Bin Mansoor

Shamyl Bin Mansoor is a technology entrepreneur, product builder and STEAM educator. He is Co-founder and CTO of LearnOBots and works on educational technology, robotics and interactive learning products. Learn more about him.

Related Articles

How to Choose an Educational Robotics Platform for Your School: A Practical Guide for Educators and Administrators

A practical guide for schools evaluating educational robotics platforms. Covers age-appropriateness, hardware vs. simulation, curriculum integration, budget realities, and what questions to ask before buying.

Sakana AI's MASS: Can a Model Teach Itself by Watching Its Own Agent Teams?

Sakana AI's MASS paper shows that a single language model can improve itself by running multi-agent teams, evaluating the results, and training on the best executions. After two cycles with Qwen3.6-27B, the model achieves 1.2-1.6x higher performance per output token on public benchmarks. Here is what the method does, where it falls short, and what it means for product teams and AI safety.

Strands Decider 2B: Why Open-Source Decision Models Are the Next AI Architecture Shift

AWS's Strands Agents released Strands Decider 2B, an open-source 2-billion-parameter decision model that runs locally on consumer hardware and costs nothing to operate after training. Here is what it means for product teams, CTOs, and developers building agentic AI workflows.