Google DeepMind Gemini Robotics vs Physical Intelligence (π)
Google DeepMind Gemini Robotics vs Physical Intelligence (π): side-by-side comparison of two robotics foundation models & agent stacks systems — architecture, taxonomy, license, pricing, MCP/A2A support, and direct edges.
Cost & capability
| Google DeepMind Gemini Robotics | Physical Intelligence (π) | |
|---|---|---|
| Capability band | entry | entry |
| Capability composite | 35 | 32 |
Where they differ (10)
Rows where both sides have data and the values disagree — the shortlist of dimensions that actually distinguish these two systems.
| Google DeepMind Gemini Robotics | Physical Intelligence (π) | |
|---|---|---|
| Capability composite | 35 | 32 |
| Type | Robotics foundation model from Google DeepMind | Robotics foundation-model lab (π0 / π0.5) |
| Created | 2025-03 | 2024-03 (founded); π0 released Oct-2024 |
| Latest release | Gemini Robotics 2 (with Gemini Robotics ER 2 and On-Device 2; date not stated on the model page) | pi0.7 (2026-04-16) — model; openpi repo has no GitHub releases |
| Funding | Google DeepMind (Alphabet parent) | $400M Series A Nov-2024 (Bezos / OpenAI / Thrive / Lux; $2.4B val) |
| Backend storage | Google Cloud | Caller manages robot trajectories |
| Deployment | Trusted-tester via partners; Vertex AI / Cloud planned | Both (self-hosted + managed cloud) — openpi weights self-hosted; Pi API beta (managed fine-tuning + hosted inference, waitlist) since 2026-10-13 |
| API surface | Gemini Robotics-ER 2 preview in Google AI Studio (gemini-robotics-er-2-preview); VLA models via trusted-tester waitlist | Python; checkpoints on HF; robot hardware integration via repo |
| Optimised for | Robotics built on Gemini 2.0 multimodal foundation | General-purpose robot foundation models — cross-embodiment manipulation |
| Anti-fit | Closed weights; trusted-tester only; Google Cloud lock-in expected | Research-stage — no productised SaaS endpoint; needs robot hardware to deploy |
At a glance
| Google DeepMind Gemini Robotics | Physical Intelligence (π) | |
|---|---|---|
| Section | Robotics foundation models & agent stacks | Robotics foundation models & agent stacks |
| Tier | T1 | T1 |
| Type | Robotics foundation model from Google DeepMind | Robotics foundation-model lab (π0 / π0.5) |
| Created | 2025-03 | 2024-03 (founded); π0 released Oct-2024 |
| Latest release | Gemini Robotics 2 (with Gemini Robotics ER 2 and On-Device 2; date not stated on the model page) | pi0.7 (2026-04-16) — model; openpi repo has no GitHub releases |
| License | — | OpenPI Apache 2.0 (Feb 2025 release of weights) |
| GitHub | — | Physical-Intelligence/openpi — ~14.4k stars, last commit 2026-08 |
| Funding | Google DeepMind (Alphabet parent) | $400M Series A Nov-2024 (Bezos / OpenAI / Thrive / Lux; $2.4B val) |
| Backend storage | Google Cloud | Caller manages robot trajectories |
| Deployment | Trusted-tester via partners; Vertex AI / Cloud planned | Both (self-hosted + managed cloud) — openpi weights self-hosted; Pi API beta (managed fine-tuning + hosted inference, waitlist) since 2026-10-13 |
| API surface | Gemini Robotics-ER 2 preview in Google AI Studio (gemini-robotics-er-2-preview); VLA models via trusted-tester waitlist | Python; checkpoints on HF; robot hardware integration via repo |
| Multi-tenancy | Per-customer (trusted-tester) | — |
| A2A | Google's own A2A protocol — likely first-party | — |
| Optimised for | Robotics built on Gemini 2.0 multimodal foundation | General-purpose robot foundation models — cross-embodiment manipulation |
| Anti-fit | Closed weights; trusted-tester only; Google Cloud lock-in expected | Research-stage — no productised SaaS endpoint; needs robot hardware to deploy |
Pros & cons
Google DeepMind Gemini Robotics
Pros: Built on Gemini 2.0 (best-in-class multimodal); DeepMind RT-X lineage; Apollo + Agile Robots partners; A2A first-party.
Cons: Closed weights; trusted-tester only; Google Cloud lock-in; no developer-self-serve.
Physical Intelligence (π)
Pros: Strongest pure-play robot FM lab; $400M Series A; OpenPI weights released; SOTA manipulation demos.
Cons: Research-stage; no commercial endpoint yet; requires hardware; cross-embodiment generalisation still polarising.