I recently visited a robotics event in Hong Kong to observe physical AI outside the usual promotional context. Two test environments stood out. The first was a retail simulation where a mobile robot maps shelves and identifies where products are located. The second was robot soccer, which creates a compact test of movement, balance, coordination, timing and recovery. My three main observations: 1. Physical AI is a systems problem. A capable model is only one layer. The robot also needs reliable perception, localization, control, actuators, power and mechanical stability. 2. Imperfect demonstrations are highly valuable. When a robot loses balance or requires operator assistance, it exposes the real engineering bottlenecks instead of hiding them. 3. Human operations remain part of the product. Calibration, maintenance, recovery and supervision are still important parts of many real-world deployments. The result is not simply “AI inside a robot.” It is an operational stack that combines intelligence, hardware and ongoing support. Which physical-AI use case do you believe has the clearest commercial path today?

New workflow experiment: I used Google Stitch as the design surface, then connected the process with Codex CLI + MCP as an agentic iteration layer. The first working UI direction came together in ~5 minutes. This is not about replacing design taste. It is about compressing exploration and turning AI generation into a repeatable workflow. If you want, reply STITCH and I’ll share the setup path. Try you own workflows from here: https://sell.systems