PhysicalAI.bestOur standard
Curated comparison · General robot policy evaluation

Gemini Robotics 2 vs Physical Intelligence π0.7

Both releases address robot behavior, but partner access and published policy resources impose different evaluation paths. Keep exact revisions, tasks and robot embodiments attached to any result.

Source contextPrimary specCompany claimReviewed
Sources 3
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Gemini Robotics model family

Google DeepMind · Publication date not disclosed · Source accessed 2026-09-28

π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

Reviewed 10 decision dimensionsNo universal winner
Version and date boundaries matter. The exact revision below defines each column. A deployment of an earlier revision is not evidence that a newer revision is in production. Missing fields remain unknown.

On small screens, swipe the table horizontally. Each row carries its own source and review date.

Decision dimensionGemini Robotics 2 Exact revision / scope
Gemini Robotics 2
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Physical Intelligence π0.7 Exact revision / scope
π0.7; arXiv:2604.15483v2
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

Developer

The developer identity does not establish a deployment relationship.

Google DeepMind
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Physical Intelligence
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

Exact version

Keep generations, checkpoints and software revisions separate.

Gemini Robotics 2
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

π0.7; arXiv:2604.15483v2
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

System or model type

Choose alternatives that address the same decision.

Vision-language-action
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Vision-language-action
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

Access conditions

Availability of a paper is distinct from access to a model or service.

Early-access partner program
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Public technical paper; public checkpoint access not verified
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

License / reuse

Code, checkpoints, datasets and assets can have different terms.

Not publicly disclosedNot publicly disclosed
Inputs

Observation and scene formats determine integration work.

Vision and language
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Language instructions, multimodal context, performance metadata and subgoal images
Primary specReviewed
Sources 1
π0.7 original research paper

Physical Intelligence · Published 2026-04-16 · Source accessed 2026-09-28

Outputs

An action policy, image generator and simulator return different outputs.

Robot motor actions
Primary specReviewed
Sources 1
Gemini Robotics 2 launch and access

Google DeepMind · Published 2026-07-30 · Source accessed 2026-09-28

Not publicly disclosed
Embodiment

Hardware and control boundaries are part of the evaluation.

Not publicly disclosedNot publicly disclosed
Compute

Confirm requirements for the intended workload.

Not publicly disclosedNot publicly disclosed
Integrations

Verify exact versions, interfaces and support responsibilities.

Not publicly disclosedNot publicly disclosed

Choose based on your constraints.

These are evaluation paths, not a ranking. Verify the current terms and run a task-specific trial before committing.

You need to plan an accessible evaluation

Inspect each release’s current access terms before designing an experiment. A paper, demo or model-family name is not a guarantee of checkpoint availability.

You need comparable performance evidence

Use the same task distribution, observation inputs and embodiment, or clearly report the mismatch. No cross-benchmark total score is meaningful here.

Evidence gaps to resolve

  • Published fields describe the identified revision only; configuration and source dates may differ.
  • No matched independent operating trial is asserted by this comparison. Missing price, reliability and intervention figures remain unknown.
  • Benchmark results require the same protocol, inputs and revision context. Cross-benchmark totals are not calculated.
Inspect the evidence

Sources & evidence

md-gemini-models
Gemini Robotics model family

Google DeepMind · Official spec · Publication date not disclosed · Source accessed 2026-09-28

Factual summary and attribution only; no upstream prose, images, weights, or dataset redistributed.

Referenced by 9 records: Elder Care · Gemini Robotics 2 · Gemini Robotics ER 2 · Gemini Robotics On-Device 2 · What Is Physical AI?

Show 4 more records

Physical AI vs Embodied AI · What Is a VLA Model? · Demo vs Pilot vs Production Deployment · Physical AI Market Map

How comparison evidence worksMore curated comparisons