Frontier Models for Every Robot: Announcing our API beta

TL;DR

Physical Intelligence opens the Pi API beta, giving partners hosted fine-tuning + inference access to the latest model in the π line (post-π0.5, π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization π0.7, π*0.6: a VLA That Learns From Experience (RECAP) π*0.6/RECAP; the next-in-line Eren Chen teases Physical Intelligence π0.8 announcement π0.8 was teased two weeks earlier). The pitch is adapt-via-API rather than train-your-own-foundation-model: upload demonstrations, launch fine-tuning on PI-managed compute, access the fine-tuned checkpoint via a cloud-hosted inference gateway. Beta pilots span strawberry harvesting through server assembly, deployed in Japan, Switzerland, and across the US, mostly running cloud inference. Blog is positioning + screenshotted customer deployments; no architecture, benchmark numbers, latency numbers, or pricing disclosed.

Key claims

The Pi API makes the “latest and most powerful robotic foundation models” (unnamed successor to π0.5; the next release per Eren Chen teases Physical Intelligence π0.8 announcement is π0.8) available for fine-tuning and inference without partners having to run ML infrastructure, compute, or data pipelines themselves [blog ¶1-2].
The latest model behind the API provides “more data-efficient fine-tuning and improved capability” than prior π releases in the partner collaborations run during the private beta [blog ¶2].
The common iteration loop across diverse deployments is identical in shape — collect demonstrations, train, test on robot, use failures to decide what data to collect next — and the API is designed to let partners run that loop themselves rather than requiring PI to run it per customer [blog ¶5-6].
Early partners “quickly got started and began making progress” at adapting the API-served model to their own tasks; applications span garment handling, kitchen ingredients (ambiguous: “an ingredient that wouldn’t stay in a scoop”), and manufacturing-part placement [blog ¶5-7].
Most early deployments ran with cloud-hosted inference through the API (not on-robot inference), with network connectivity flagged as the main surfaced limitation [blog ¶8].
PI will keep updating the underlying model so API users inherit new capabilities over time without changing their integration [blog ¶2].

Method

Not a technical paper — a product-launch blog post describing an API offering. Architecture is reused from the base π line already filed under VLA Models (VLM backbone + flow-matching action expert in the π0/π0.5 lineage, extended through RECAP-style advantage conditioning in π*0.6: a VLA That Learns From Experience (RECAP) and steerable composition in π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization). The specific model served through the API is unnamed in the post; the teased Eren Chen teases Physical Intelligence π0.8 announcement π0.8 is the most plausible candidate on cadence but PI has not confirmed. Fine-tuning recipe not disclosed; data-upload format, action-space abstraction, and the fine-tuning compute envelope are not described.

Results

No quantitative results reported. Qualitative evidence is screenshotted partner deployments (not reproduced here) covering strawberry harvesting, server assembly, laundry handling (inside-out garment task), kitchen prep (scooping task), and manufacturing part insertion. Geographic distribution noted: Japan, Switzerland, and multiple US sites. Only reported limitation is network-connectivity bumps for some of the cloud-inference deployments.

Why it’s interesting

This is the first filed artifact where a specific robotic-foundation-model lab commits publicly to the deployment-company model vedantnair’s quoted tweet names — PI provides the frontier model, partners build/assemble their own hardware, fine-tune on vertical-specific data via the API, and own the deployment integration. Three things worth pulling out:

(1) Operational instantiation of the deployment-economics thread. The post reads like PI’s answer to York Yang (Dyna Robotics) — ROI and scalability are inseparable; every deployment must make the next one better York Yang’s “every deployment must make the next one better” claim — the API is the mechanism PI is betting on to collect the “common pattern” across customer deployments (collect → train → test → decide next data) so that each adaptation feeds back into “the next generation of models”. Sibling in framing to Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics’ Roboport (teleop → DAgger → scaled deployment on a vertically integrated stack) and MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory’s 99.9%-via-cheap-DAgger thesis, but with the opposite integration bet: Robo Robotics / MicroFactory own hardware + policy + ops, PI owns only the model and gives partners the fine-tuning plumbing. Horizontally splits the deployment stack where those entries vertically unify it.

(2) Where this lands on the VLA Models recipe-lever board. The API itself is not a new recipe lever — it serves one of the π-line models whose recipes are already catalogued (action-pretraining-at-scale, RECAP RL on flow-matching, steerable composition). The novelty is on the deployment-pipeline axis, parallel to the runtime-plane entries (APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin APXinf-robo, Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots Embodied.cpp, InstinctFlash — High-Performance Serving Runtime for Robotics Models InstinctFlash) that let third parties serve π0.5 on their own hardware. PI’s choice is cloud-hosted inference by default (vs the Jetson-local line those runtimes target), with the trade-off visible in the post itself — the limitation flagged is network connectivity, which is exactly what on-device-first runtime stacks solve for.

(3) What’s not in the post. No model name, no architecture deltas vs π0.7 / π*0.6, no fine-tuning compute or data-size numbers per partner, no latency or throughput numbers for cloud inference, no pricing. Readers tracking the π-line recipe board should wait for the π0.8 companion report (per Eren Chen teases Physical Intelligence π0.8 announcement) or an API-specific technical post before updating positions on VLA Models.