What is a robot foundation model?
A robot foundation model is a large AI model pretrained on broad robot, simulation and web data, then adapted to many tasks and robot bodies instead of being trained for one job.
Updated:
Definition
A robot foundation model is a large neural network trained on broad and varied data so that one model can be adapted to many tasks, environments and robot bodies. The idea comes from language models: pretrain once at scale, then specialise cheaply.
How it works
- Pretrain on a mix of robot trajectories, simulation, human video and image-text data, often starting from a pretrained vision-language model.
- Post-train on demonstrations from the target robot and tasks.
- Deploy, usually with a fast low-level controller underneath that turns the model's output into motor commands.
The output can be actions (a VLA model), predicted future states (a world model) or a plan in language that other modules execute.
Why it matters
Classic robot software needs a new program or policy for every task. Foundation models aim to make new skills a matter of data and prompting instead. Their bottleneck is data: robot experience is scarce compared with web text, which is why teams invest in teleoperation fleets, motion capture and synthetic data. See also embodied AI.
Frequently asked questions
Is a VLA model a robot foundation model?
Yes. A VLA is one kind of robot foundation model, the kind that outputs robot actions directly. Others predict future video (world models) or plan steps in language.
Where does the training data come from?
Teleoperated robot demonstrations, datasets pooled from many robot types, simulation and synthetic data, human video and motion capture, plus the web knowledge inherited from the vision-language model it starts from.
Does one model work on any robot?
Not out of the box. Most need fine-tuning with data from the target robot, although training across many robot bodies reduces how much new data that takes.
Coverage
NEURA's Neuron AI adds touch, sound and memory to the robot foundation model
NEURA Robotics launched Neuron AI, a 'perception-language-action' model that fuses 3D vision, touch, force, audio, robot state and language with memory, and will demo an 11-minute laundry task on its 4NE1 humanoid.
Industrial Next says its stations retool in minutes; BMW will evaluate them
At TechCrunch Disrupt, the San Francisco startup said foundation-model workcells already peel film and seat connectors on a major phone line, and that BMW's technology office will test the approach.
Physical Intelligence opens its π0 robot models to everyone via an API
The Pi API beta lets robot builders upload data, fine-tune π0/π0.5 on Pi's infrastructure and run the result through hosted inference.
Rhoda's robot finds textured balls by touch, with its cameras blocked
Rhoda AI's Oct. 13 post pairs a new tactile sensor with a model that predicts future touch. In a blindfolded demo, the robot retrieved textured balls it had not trained on.
XPENG sets IRON humanoid deliveries for the first half of 2027
At the Paris Automotive Summit, XPENG said IRON targets mass production in late 2026, first deliveries in H1 2027 and a European rollout in H2 2027.
Mecka raises $60M Series B to digitize human motion for robot training
Sequoia led Mecka's $60M Series B with NVIDIA, Microsoft's M12, Qualcomm Ventures and Samsung; the startup captures human demonstration data and says it passed $100M run-rate revenue in June 2026.
Skild's S1 learns soccer by playing itself for 140 simulated years
After launching S1 — a robotics foundation model that copies a single video demo without fine-tuning — Skild showed Physical Self-Play: the same base policy trained against itself in Isaac Sim, then transferred to a humanoid.
Figure's Helix 2.5 cleans 30 homes it has never seen
Figure says its humanoid tidied rooms, folded towels and made beds in 30 unseen Bay Area homes with zero local data, backed by a $3.5B compute deal with Nscale.
Companies
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.