Models & learning
Robot foundation model
A robot foundation model is a broadly pretrained model intended to provide a reusable starting point for multiple robot tasks, environments or embodiments. It learns from diverse robotics and sometimes web or human data, then acts directly or is adapted with target-domain data. The term describes a training and reuse strategy, not one fixed architecture.
Also known as: RFM, robotics foundation model, robotic foundation model
Updated
What makes the model foundational
A robot foundation model is pretrained across enough variation to be reused rather than rebuilt for one fixed task. Reuse may mean direct control, fine-tuning, adding a new action head or supplying the model to a planner. “Foundation” describes that broad pretraining and downstream adaptation pattern; it does not guarantee general-purpose autonomy.
Implementations include vision-language-action models and diffusion policies. A VLA is defined by its visual, language and action interface, whereas a robot foundation model is defined more broadly by the intended scope and reuse of its pretraining.
Why diverse robot data matters
Open X-Embodiment pooled datasets from different institutions and embodiments in a common format, then trained RT-X policies to study transfer between robots. Its data still contained different cameras, observations and action spaces. Standardising records makes joint training possible; it does not make the bodies interchangeable.
Octo is a concrete reusable policy. It was pretrained on a mixture of Open X-Embodiment trajectories and designed for fine-tuning with new sensory inputs and action spaces. This illustrates why dataset metadata and embodiment-specific transforms are part of foundation-model training, not clerical details.
Application to humanoids
A manipulation model can provide useful visual or task representations without already controlling a humanoid. Whole-body balance, bimanual coordination, joint limits and control rates create a different action domain. Adaptation data should therefore identify the humanoid’s observation schema, command representation and safety constraints. Evaluation must also test sustained physical execution and recovery, rather than treating pretraining breadth as evidence that every supported-looking task is reliable.
Sources
Related terms
Models & learning
Vision-language-action model
A vision-language-action model is a robot policy that conditions on visual observations and natural-language instructions to predict executable actions. It adapts a pretrained vision-language model, or a similar multimodal backbone, using robot trajectories so that semantic knowledge is connected to an embodiment’s action space. The action representation may use discrete tokens or continuous values.
Data & collection
Cross-embodiment data
Cross-embodiment data is robot training data drawn from multiple physical embodiments, such as arms, mobile manipulators, quadrupeds or humanoids with different kinematics, sensors and action spaces. The datasets are aligned or packaged so models can learn jointly from experience produced by different robots, although shared formatting does not make their observations or controls physically equivalent.
Models & learning
Imitation learning
Imitation learning is a family of methods that learns a policy from examples of expert behaviour rather than specifying every control rule by hand. In robotics, demonstrations pair observations or states with actions, trajectories or inferred objectives. Behaviour cloning is one imitation-learning method; interactive and inverse approaches address different supervision and distribution-shift problems.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.