What to watch, read and learn in robotics and physical AI, curated from the 2026 posts of Michelle Sun (@michellelsun), investor across the physical AI value chain and author of the Core Matter research newsletter. Every item links back to the post where she shared it.
If you come from software: start with the map, then data and models, then the hardware that limits them, then the business side.
Robot data isn't generated by accident; every hour is paid for. She maps teleop, egocentric video, UMI and simulation on fidelity, diversity and scale, and asks whether a "Scale AI of robotics" can make margins.
From Jim Fan's "RIP VLAs" to 1X predicting world models win by 2027, Dyna-2, Odyssey-3 and Gemini Real2Sim. She also flags the counterpoint paper that both miss the grounding problem.
Hands unlock economic value, legs get attention. Hand architectures, tactile sensing, price spread from $104 to $100k, and IROS converging on contact-rich manipulation.
Her most-liked threads: actuator bills of materials, supplier maps from mine to joint, China's supply base, rare-earth magnets, and the EV/robot supply overlap.
Unitree and AGIBOT shipped ~5,000 humanoids each in 2025 vs ~150 for Figure and Agility; a wave of Chinese IPOs; FCC Covered List and tariffs on the US side.
"The demo is 1% of the business." Fixed scenes, single tasks, environment failures, and a growing market for installing and running robots.
NVIDIA Halos, Safeworld, Agility's functional safety, and the open question of who builds the attestation layer.