Robots have been "almost there" since the 1960s

Robots have been "almost there" since the 1960s.

Yesterday at AUTONOMOUS, it felt like "there" is finally within reach, with industry veterans voicing their excitement that a turning point for robotics is imminent.

The energy was inspiring: more startups building robots, more software and AI companies building for robotics, and a shared sense that the problem space is large enough for new entrants to gradually grow into. We've seen a healthy ecosystem grow this way in the LLM space, and there's a real opportunity for robotics to follow suit.

A few things surprised me.

There were multiple mentions of a bias toward world models. The worry with VLAs is that they break out of distribution. One bad action creates an unfamiliar observation, which creates a worse action, and the policy death-spirals from there.

This doesn't mean end-to-end learning is irrelevant, or that VLAs are dead. The broader view was that, for now, an end-to-end policy will live inside a larger modular system rather than run as a pure VLA.

There is a classic precedent for this in jidoka, one of the ideas behind the Toyota Production System. The concept grew out of automatic looms built to stop themselves when a thread broke; later, andon cords let workers halt the production line when they saw a problem.

The same principle shows up in robots today: physics-based controllers to bound behavior, robots that ask for help when unsure, and a world model folded in as another layer of grounding.

Among the teams building end-to-end autonomy, many expected to lean harder on RL and simulation. In practice, imitation learning gets you much further than most expected. Now teams are rushing to build high-quality datasets that unlock new emergent capabilities they can test in the field.

The hardware preferences were pragmatic. Hands are an exciting technological advancement, but for deployment, teams still prefer simple grippers and bimanual arms on wheeled mobile bases.

The case for a wheeled base isn't only simplicity but also safety. We already know how to move mobile bases around safely, and that knowledge transfers directly.

We're excited to do our part to put more robots in the world.