Robot Foundation Models

Robot Foundation Models
General-purpose learned models applied to robot tasks

Robot foundation models is a research-oriented term for broadly trained models intended to support multiple robotic tasks or environments. Some models combine image and language representations with actions. Different projects use different training data, action interfaces, architectures, and evaluation methods.

Vision–language–action systems

Google DeepMind described RT-2 in 2023 as a vision-language-action system trained using web and robotics data. It demonstrates one approach to representing robotic control alongside visual and linguistic information.

Training and evaluation

Training may use robot demonstrations, simulation, teleoperation records, visual-language material or other datasets. Evaluation should report task definitions, operating conditions, baselines, failure modes and whether results transfer across robots.

Deployment limitations

A capable research model may still fail in situations unlike its training and tests. Physical deployment requires constraints on actuation, authentication of commands, monitoring and risk assessment. The term "foundation" does not provide a safety certification.

Further reading

• AI Risk Management Framework — NIST

See also

• Embodied AI
• Computer Vision
• Machine Learning
• Physical AI
• Autonomous Systems
• Embodied AI