Vision-led robotics reflects data economics

Vision first robotics today is mostly a supply chain story, not proof that vision alone closes the manipulation problem. Cameras are cheap enough to put on every robot and easy enough for small teams to stream into training loops, while force and tactile hardware still arrives as industrial gear built for expensive factory integrations. That pushes labs toward pixel to action models because the easiest data to collect becomes the default training recipe.1 2 3 4 5

• Google DeepMind RT-2 and Gemini Robotics, Physical Intelligence π0, and NVIDIA GR00T are all framed around vision language action stacks. Their common thread is not that touch is solved away, but that web scale image and video data already exists, while robot force data still has to be collected the hard way on real machines.2 3 4 5
• In practice, force and tactile data are expensive twice over. The sensor itself can cost thousands of dollars, then a team still has to mount it, calibrate it, and keep it working across many robots. That matters because most physical AI teams are small, so every extra sensor can turn a 20 robot data fleet into a capital and integration headache.1
• The strategic divide is emerging between model labs that train on abundant camera data, and hardware stacks that try to unlock the missing senses. 1X already combines cameras with tactile joint sensing, and Generalist is exposed to the same pressure, because the last jump from decent demos to near perfect task completion depends on feeling slip, pressure, and contact, not just seeing the scene.6 7 8

The next phase of physical AI will likely be won by whoever makes non visual data cheap and standard, not by whoever adds the most cameras. As force and tactile sensing move from custom industrial parts into repeatable developer kits, model design will shift from pure pixel to action systems toward multi sense stacks that can handle real world contact at production reliability.1 3 5 6