Two-week catch-up: an interface fortnight, not a robot fortnight

This is a two-week catch-up covering 10–23 October 2026; the desk was dark from the 10th. It was an interface fortnight, not a robot fortnight. Mitsubishi Electric put a humanoid inside the NC system, FANUC made machine-tool signals plug-and-play for AI, and Yamazaki Mazak picked an AI partner to write process plans. Nobody shipped a humanoid onto a CNC, and the record order book behind all of it is datacentre money.

What landed

15 Oct — The humanoid line with numbers, and the number that matters is seconds

Leju Robotics and Zhaofeng presented 15 Kuafu humanoids in routine operation on a parts-loading line at Zhaofeng’s No 3 plant in Xiaoshan. The figures are company-reported: 15 stations, more than 10 part types, parts of 80–120 mm inner diameter, 98% loading success, about 1,000 hours of continuous running and more than 60,000 parts.

The comparison that decides the item is the cycle. It is under 40 seconds, against roughly 15 seconds for a skilled worker at the same station. The Zhaofeng chairman said fixed arms already do some of these steps in 10–15 seconds, and that a humanoid needs about 15 seconds before wider rollout, which he expected in six months to a year. This is loading, not machining.

16 Oct — Mazak picks a Japanese AI partner for the process plan itself

ABEJA said it has begun joint development with Yamazaki Mazak on three things: agents that draft machining process designs (method, sequence, cutting conditions), agents that build and revise production schedules, and a robot behaviour-learning environment running from data collection through to verification on real machines.

The stated destination is Mazak machine tools and automation systems, with a human in the loop. “In the future” is the operative phrase. No terms, schedule or customers.

21 Oct — Mitsubishi Electric defines the humanoid as a supported peripheral of the CNC

Release No 3946 describes a machining system with three parts: an AI task planner that takes an operator’s plain-language instruction and assigns work between the CNC and a humanoid, a CNC with physical AI that exchanges operational information in natural language, and a general-purpose humanoid. Named tasks are the ones that resist fixed automation: setup, cleaning, removal of finished parts, inspection and troubleshooting.

Mitsubishi’s argument is that the hard part was never the robot but the per-robot communication and control sequences. Putting coordination in the NC system removes the need to integrate each robot separately. A demonstration is scheduled at JIMTOF from 26 October. The “industry first” is footnoted to Mitsubishi’s own research as of 21 October and scoped to NC systems built around humanoid collaboration.

22 Oct — FANUC makes machine-tool data plug-and-play, and adds Copilot

Two FANUC releases landed the same day. The first pre-configures machine-tool-builder-defined signals in the FIELD system Basic Package. Select a machine and you get production performance, defect information, tool-change status, tool life, robot waiting time and pallet-change waiting time, without hunting for signal addresses that mean different things on each machine. The point is mixed fleets: factory-wide analysis across several builders.

The second combines FsBP shop-floor data with Microsoft 365 business data through Microsoft Copilot, and explores Microsoft 365 Work IQ, so both can be queried in natural language. No product name, price or date.

23 Oct — The order book says datacentres, not humanoids

JMTBA’s September figures: ¥223.21bn in machine-tool orders, up 60.4% year on year, a record since the statistics began in 1988 and above the previous peak of ¥203.38bn in June 2026. Foreign orders were a record ¥159.28bn; domestic orders were ¥63.93bn.

At the 22 October press conference, chairman Shigetomo Sakamoto called the level unprecedented and said the risk to watch is a pause in data-centre and semiconductor demand. JMTBA and Nikkei credit AI-related demand, alongside job shops and aerospace. Nobody credits humanoids.

Gaps / contradictions

• Mitsubishi’s “industry first” is vendor-scoped: its own research, NC systems built around humanoid collaboration, as of 21 October. No humanoid maker, customer, price or ship date is named.
• Mazak does not have physical AI in its machines yet. ABEJA’s release says the agents will be implemented in the future.
• Leju’s 98% loading success is a reliability result, not a productivity result: the cycle is under 40 seconds against about 15 for the person beside it.
• No source discloses recovery time after a failed grasp, or whether a person still stands at the Zhaofeng station.
• No one in this window named a customer for a humanoid on a CNC, published an installed machining cell, or timed a humanoid tending cycle against a human.
• Nikkei reported FANUC’s voice-instructed AI vision for machining on 22 October, shipping January 2027. FANUC’s own page and listing date from 27 October, so we hold the formal launch for next issue.
• The September order record is not robot demand. The over-investment warning sits in the same story as the record.

What to watch

• Whether Mitsubishi’s JIMTOF demonstration names a humanoid partner or a customer site, or stays a booth demo.
• Whether ABEJA and Mazak produce a dated product, a machine or cell shipping with the agents, rather than a platform story.
• Whether anyone publishes a timed humanoid-to-CNC tending cycle against a human.
• FANUC’s formal launch of the vision system, next issue.
• The JMTBA October preliminary figures, due 11 November, as the first test of the chairman’s datacentre-pause worry.