Carbon Robotics unveils Large Plant Model trained on 150 million plant images

The article places the Large Plant Model in a broader pattern: agricultural robots are advancing alongside foundation-model AI. Rather than building a model for each crop, the approach is to pre-train broadly and then fine-tune minimally for each field, and Carbon Robotics applies that idea to plants. That matters because fields vary widely, and the speed of adaptation largely decides how fast harvesting and weeding robots become practical.

Underneath it all sits a labor problem. Japan's core agricultural workers fell to 1.036 million in 2025, down 56.8% from 2.4 million in 2000, with an average age of 67.7 and 69.6% aged 65 or older. Robotics players have emerged both domestically (AGRIST's cucumber harvester Q, Kubota's autonomous combine with NVIDIA) and abroad (eternal.ag's greenhouse Harvester, which raised 800万ユーロ, and Agroz with UBTECH's Walker S humanoid).

Because farms are outdoors and uneven, the article notes that robotics here is technically as hard as construction sites, so lowering upfront cost through rent or Robot-as-a-Service is a common route — AGRIST's planned 2026 rental fits that. Japan's 2024 smart agriculture law backs this with priority goals and tax and financing support. Whether any of it pays off is likely to hinge on how quickly models adapt to each field and whether Level 3 rules, including liability in accidents, get resolved.