Sources
Every figure in the "why robotics, why now" story on the homepage traces back to one of these. Anything approximate, or still pending a primary citation, is flagged below rather than smoothed over.
AI got 300× cheaper. Building didn’t.
Published API price sheets, GPT-4-class or better, $ per 1M input tokens. Building: US construction productivity, flat for decades (Richmond Fed 2025).
Physical AI is outpaced by digital AI.
Release dates: each vendor’s own announcement. Lead time: 128-week average, Wood Mackenzie Q2 2025 T&D supply-chain survey. Which models count as frontier is a judgement call.
The models aren’t good enough.
Best real-robot success rate per model on RoboDojo, RoboChallenge and PhAIL (pick and place), as listed on b2atoms.com/data/leaderboards. Production lines expect 95–99%.
Simulation scores don’t survive a real robot.
LIBERO simulation scores as run by BAAI (FlagEval-Robo), against the same model’s best real-robot success rate on the leaderboards above.
There’s no easy way to scale robot training.
Robot models learn from simulation or from people driving a robot by hand (teleoperation). The simulation gap is the previous slide; teleoperation is how the large open datasets were collected, e.g. Open X-Embodiment (2023) and DROID (2024).
The robot is the cheap part.
Consistent across independent integrators: the arm is 30–50% of project cost, integration and programming 20–50%. The split beyond the arm is illustrative.