Doosan Robotics targets a 2028 humanoid, working up from the arm

Doosan Robotics chief executive Kim Min-pyo set out a three-stage path to industrial humanoids at the 2026 Global Machinery Technology Forum on September 7, reported September 8, 2026: collaborative robots first, intelligent robot solutions second, industrial humanoids third. The form factor widens in step, from single-arm to dual-arm to mobile platforms, with a humanoid proof-of-concept targeted for 2028 in skilled tasks such as surface treatment and welding.

The sequencing argument is the substance of it. “It is not whether to develop humanoids, but in what sequence,” Kim said, describing an approach that establishes task intelligence first and settles the physical form afterwards — the reverse of building a general-purpose body and then looking for something for it to do.

Up from the arm, not down to the hand

Doosan is approaching the humanoid from the opposite end to almost everyone else in the category. A cobot maker already owns the hard parts nobody demos: force and torque control, safety certification for shared workspaces, and years of logged contact data from real customer sites. What it lacks is locomotion and whole-body coordination. Humanoid startups own the walking and lack the manipulation credibility — which is why so many of their pilots stall at the point where the robot has to touch something under load.

Kim was explicit that physical AI is a distinct problem from language modelling: it needs real-world data on contact, force, friction and failure, not internet text. That is a defensible reason to start where the data already exists. The counter-argument is that manipulation skill has never been the thing investors underwrite — general-purpose morphology is — and arriving with excellent welding and no legs is a harder story to fund than the reverse.

Agentic robots as the middle step

The second stage carries most of the technical risk. Doosan’s framing of an “agentic robot” is one that takes only the objective, then interprets the environment, plans, executes and self-corrects — against today’s industrial norm of an engineer specifying every motion. The company has been building toward this since April 2026, when it announced work with NVIDIA on simulation and learning infrastructure around its own agentic robot operating system, with intelligent robot solutions slated for 2027 and industrial humanoids commercialising from 2028.

Running the first proof-of-concept inside Doosan group factories is a real advantage. Gas turbine machining and heavy hydraulic assembly are captive, high-torque, high-tolerance processes with a customer that cannot walk away — the kind of deployment that produces failure data rather than a press release.

The date is the problem

2028 is late measured against Korea’s own procurement clock. The government plans to buy 250 domestically made humanoids in 2027 and 1,080 by 2030, inside a ₩2.3 trillion full-stack programme announced in August 2026. A proof-of-concept starting in 2028 puts Doosan outside the first procurement cycle and possibly the second, in a market where the state is the anchor buyer and early wins set the reference deployments everyone else is compared against. Meanwhile Chinese vendors took more than 97% of roughly 19,100 humanoids shipped globally in the first half of 2026. Doosan’s route may be the sounder engineering path. It is still arriving after the tender. The hedge is that Doosan does not need the humanoid to sell: stages one and two are shipping products, and if the third stage slips, the company is still a cobot vendor with an AI layer rather than a startup with nothing to invoice.