What Is a Robot Foundation Model?

Look for the reusable starting point

The useful meaning of foundation is that a learned system can support more than one narrowly specified downstream task. That does not promise universal competence. In a review, record what was pretrained, what was adapted and what was evaluated after adaptation. GEN-1’s description separates human-activity pretraining from robot-task adaptation, illustrating why those stages should not be compressed into one claim of generality.

Separate kinds of generalization

A new instruction, object, room, task and robot body each change a different part of the problem. Ask which changes an experiment actually made and which stayed fixed. Skild’s company description discusses multiple robot forms; this is a vendor claim that needs embodiment-specific evaluation. A model can perform well on one kind of variation while leaving another untested.

Trace training material and adaptation

Record the source datasets, collection setup, data mixture and any task-specific training when disclosed. Open X-Embodiment combines constituent datasets with their own provenance; DROID documents a real-robot collection setup. A large data label does not establish that the target site, objects or recovery conditions are represented. Ask what data must be collected locally and who can use or retain it.

Include the runtime in the model identity

An operational system may combine a checkpoint with prompting, planning, normalization, action processing and execution software. The π0.7 paper’s context-conditioning approach is an example of behavior depending on more than a family name. Record the delivered model revision and runtime configuration together. Otherwise, a comparison may attribute differences in the surrounding system to the model alone.

Demand an adaptation and acceptance plan

Before choosing a platform, ask for the target embodiment, required training examples, ownership of adaptation work and a test on held-out operating conditions. Identify what can be downloaded, what is available only through a partner and what remains a paper. A research release can be valuable evidence while still leaving support terms, production deployment and failure recovery unverified.

Additional disclosed fields

Training-source example

GEN-1 distinguishes human-activity pretraining and robot-task adaptation

Company claimReviewed 2026-09-28

Sources 1

Introducing GEN-1

Generalist AI · Published 2026-04-02 · Source accessed 2026-09-28

Context-conditioning example

The π0.7 paper describes multimodal context used to steer behavior

Primary specReviewed 2026-09-28

Sources 1

π0.7 original research paper

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

Data provenance example

Open X-Embodiment retains constituent dataset information

Primary specReviewed 2026-09-28

Sources 1

google-deepmind/open_x_embodiment official repository

google-deepmind · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Relationships & deployments

Related reading is an editorial crosslink. Sourced connections describe relationships reported in the cited material. A link to a versioned profile does not establish compatibility with that version unless the connection note explicitly identifies it.

Record history & verified changes

A research review records when we checked a source. It does not mark a product launch or a new deployment.

Initial reviewed record. No subsequent field change has been recorded.