Robots Are Finally Starting to Work

Physical Intelligence's mission is audacious: create a foundation model that can control any robot to perform any physically feasible task at a performance level useful to people in all walks of life. Co-founder Kuang Vang frames this not as a single «ChatGPT moment» but as an incremental peeling of the onion — starting with a strong base model imbued with common-sense knowledge, deploying mixed-autonomy systems (like self-driving cars today), and iteratively improving through exposure to real-world edge cases. One day, the system wakes up fully autonomous.

Historically, robotics has been a graveyard of broken promises because it demands mastery of three pillars: semantics (understanding the world), planning (deciding what to do), and real-time control (executing actions in a changing environment). Language models unlocked semantics; papers like SayCan, RT-2, and PaLM-E brought common sense into robots by adapting vision-language models to «speak robot language.» The breakthrough came with cross-embodiment training: the Open X-Embodiment dataset trained a single model on 10+ robot platforms, and the generalist outperformed specialists by 50 percent. That result shattered the old paradigm that each robot needed bespoke engineering.

Physical Intelligence now deploys models hosted entirely in the cloud, querying API endpoints in real time during high-frequency control loops. By burying inference time within the robot's execution cycle and using «real-time chunking» (pre-computing overlapping action sequences), the team eliminated the need for expensive on-robot compute. Kuang admits he has never seen some partner robots in person and intentionally avoids learning their internal systems — proving the model can parachute into any hardware stack and just work.

Robotics historically failed because it demanded full vertical integration: custom hardware, proprietary autonomy stacks, safety certifications, and direct customer relationships. The upfront capital and expertise required kept the industry in a «mainframe era» — only giant enterprises could afford deployment. Data was the bottleneck: there was no «internet of robot data» to train on, and even within a single lab, no two robots were identical due to hardware drift, software updates, and component variations.

Physical Intelligence's insight was to treat hardware heterogeneity as a feature, not a bug. By training cross-embodiment models from day one, the team discovered emergent abstractions: the model learned how to control «a robot,» not just «this specific robot.» That meant old data from drifted hardware remained useful, and new robot types could be onboarded without retraining from scratch. Cloud-hosted inference further decoupled model development from hardware constraints, letting founders iterate on software without worrying about obsolete compute units on每个robot.

The result is a new equation for startup founders: you no longer need a PhD in robotics, a custom-built robot, or years of classical-control engineering. You need workflow understanding, scrappy hardware procurement, data-collection rigor, and the ability to run mixed-autonomy trials. Physical Intelligence provides the foundation model; vertical startups provide domain expertise and go-to-market. Kuang explicitly framed this as enabling a «Cambrian explosion» — not hyperbole, but a measured prediction that hundreds of robotics companies will emerge in the next few years.