Race to Supply Robot Brains

This race is really about who sets the default benchmark for what a robot brain should already be able to do before any customer starts an integration. Physical Intelligence and Generalist are both selling a hardware neutral control layer, so every visible jump in dexterity, speed, or transfer across robot types changes what buyers expect in pilots, pricing, and deployment timelines. That makes model launches function like product launches, not just research milestones.

• Physical Intelligence has pushed this category into a faster release cadence by publishing π0 in October 2024, open sourcing code and weights in February 2025, and positioning the model as something developers can fine tune on different robots and tasks. That raises the floor for what a rival must show to look state of the art.
  2 sacra 3 physicalintelligence
• Skild is chasing the same any robot, any task vision, but with distribution as the wedge. Its March 19, 2026 partnerships with ABB Robotics, Universal Robots, and MiR, plus work with NVIDIA and Foxconn, give it a path to plug the model into existing industrial fleets and gather real world manipulation data faster.
  4 sacra 5 skild 6 skild
• The practical bottleneck is not just model quality, it is data from messy physical deployments. Across home and industrial robotics, early rollouts are being used to harvest teleoperation logs, human demonstrations, and task traces that improve the next model version, which is why one strong deployment network can matter as much as one strong paper.
  1 sacra 4 sacra 6 skild

Going forward, the winner is likely to be the company that combines the best transfer learning with the fastest field data loop. As more OEMs and developers expect one model to work across arms, mobile manipulators, and new tasks with little retraining, each release will keep redefining the minimum acceptable intelligence for the whole category.