Robotics foundation model: a new step toward the "GPT of manipulation"
In one sentence A robotics lab (Physical Intelligence or peer) publishes a new multi-embodiment foundation model for general manipulation, trained on cross-robot datasets.
In 2024-2025 AI robotics went through two parallel revolutions: humanoids (Figure, 1X, Tesla Optimus, Unitree) and robotic foundation models (Physical Intelligence's π0, Google DeepMind's RT-X). The idea mirrors ChatGPT for language: train a general model on lots of data from many different robots, then adapt it to specific tasks with little fine-tuning.
In 2026 a new step in that direction: a more general foundation model than its predecessors, able to transfer skills (grasping, folding, pouring, opening) across robots (industrial arms, bipedal humanoids, mobile manipulators) with few-shot adaptation.
For people in the sector: robotic development cadence accelerates. You no longer train a robot from scratch for every task — you start from the foundation model, fine-tune with minutes or hours of teleoperation, and it works.
For the general public: still no useful home robot, but warehouses, logistics, and manufacturing see their first real deployments with these models.
Companies
Physical Intelligence, Figure, 1X
Tools
robotics foundation model
Tags
Sources