Toyota wants 400,000 factory robots trained by its own workers
Toyota is rebuilding its factories around robots it built itself, and the teaching method is its own workforce. China's chip makers spent the same week arguing that the model belongs on the device, not in a data centre.
Toyota wants 400,000 factory robots — and its own workers are the training set.
The Toyota Motor group plans to put 400,000 in-house-developed robots into production sites worldwide as its factories are rebuilt, Nikkei Asia reported, with the automaker pitching the result as factories where people and machines work side by side rather than one replacing the other. The number is an estimate tied to renovation timelines, not a signed order — and the more interesting part is how the robots are being taught.
The robot is ELEY, short for Embodied Learning Robot for Enhanced Yield: a 50-kilogram wheeled platform with two-fingered grippers that deliberately does not walk. Toyota's engineers chose wheels after a decade of hard lessons with ELEY's predecessor, the Human Support Robot, and sized the joints to the average dimensions of an adult Japanese male — not for looks, but because robots trained by watching humans inherit systematic errors when their bodies are proportioned differently. Workers teach it by wearing jigs modeled on the robot's own fingers to demonstrate tasks that need precise hand movement, per Ars Technica.
The learning stack comes from the Toyota Research Institute: Large Behavior Models built on diffusion policy, a technique TRI developed with Columbia and MIT that treats action generation as a denoising process, so the robot can represent several valid ways to do a job instead of one fixed motion. Toyota's plan is a shared training network — a skill mastered in Nagoya propagates to a plant in Kentucky without two weeks of repeat practice. Executive vice president Hiroki Nakajima framed the goal as "a world where robots coexist with humans, rather than replacing them," which matters at a company with 18,000 veteran workers across 60 factories, including 11 in the US.
Why it matters: Toyota's stated motive is tacit knowledge. Its takumi — master craftspeople whose skills take decades to build — are nearing retirement, and Japan's shrinking labor force cannot replace them at the rate needed. Capturing embodied skill by observation is a genuine answer to that problem. It is also the part Toyota has not solved: its own engineers flag long-shift reliability, end-effector repeatability, and the data plumbing for fleet-wide learning as open gaps, and diffusion policies are known to fail unpredictably when reality drifts from training conditions. Critics will note that "coexist" is an aim, not a contract.
Houmo says its next edge chip will stack memory and compute to run 100B models at 10 watts.
At the 2026 Global AI Chip Summit in Shanghai, China's Houmo AI confirmed its next-generation edge large-model chip will use a 3D compute-in-memory architecture — fusing compute-in-memory with 3D DRAM — after six years of iterating on SRAM-based compute-in-memory designs. Zhidx reported the next-generation flagship, A20, has already taped out, and that the new architecture can place a 100 billion-parameter model on an edge chip at roughly 10 watts, with enough headroom to potentially fit inside a phone.
That target is the whole ballgame for on-device AI: compute-in-memory keeps weights next to the arithmetic that uses them, sidestepping the memory wall that makes running large models locally expensive. Houmo's current M50 runs 30B to 120B models at 10 watts with 160 TOPS and 48 GB of memory, and the company says it has passed 150 design wins in about six months of mass production.
The caveat: Houmo's own AI systems chief engineer told the summit that 3D stacking does raise memory bandwidth for real-time physical-AI workloads, but the supply chain for it is not mature yet. A tape-out is a proof point, not a product — the 100B-at-10W claim is the number to hold the company to.
What to watch: whether Toyota's first ELEY deployments stay on the demo floor or reach shift-length reliability, and whether Houmo ships A20 before the 3D DRAM supply chain catches up.
Which bet ages better — teaching robots by hand, or putting a 100-billion-parameter model in your pocket? Tell us in the comments.
Sources: Nikkei Asia · Ars Technica · Automotive World · Tech Times · Toyota Frontier Research Center · ITHome · Zhidx
