Force and Tactile Data Bottleneck

The real bottleneck has shifted from model architecture to data that captures contact with the world. Vision only gets a robot to the point where it can usually see what to do, much like early LLMs could usually produce plausible text, but production use needs the last jump in reliability, where gripping the right force, feeling slip, and recovering from small mistakes matter more than another camera frame.1 2 3 4 5

• The current VLA wave is built around whatever is cheap to collect. Google DeepMind's RT-2 and later Gemini Robotics both use vision, language, and action as the core loop, which fits a world where cameras are abundant and easy to wire into robots, while force and tactile sensing remain scarce and expensive.1 2 3
• That is why the market still looks pre breakout. Anvil describes teams spending five to six months just assembling arms, cameras, compute, and firmware before they can even start collecting training data. Foundation frames the same post 2022 shift as moving robotics from a hardware problem to a data acquisition problem.4 5
• The competitive split is already visible. Humanoid companies like Figure, Apptronik, and Agility are betting a human shaped body lets them drop into existing factories, while Anvil is betting many valuable tasks stay at fixed stations where cheaper non legged systems can gather task data faster and with less mechanical overhead.4 6 7

The next phase is a market collapse around the sensing stack that most reliably pushes task success toward 99.99%. Once force and tactile data become cheaper to collect and easier to train on, robotics will look less like broad experimentation and more like the post GPT-3.5 period, where a few winning data recipes rapidly set the industry standard.4 5