TOKYO, 31 AUG 2026 — Japan's government and TEPCO are expected to launch a consortium this fiscal year to bring AI robots and drones to the Fukushima Daiichi cleanup. Three reactors hold 880 tons of fuel debris. Two trial retrievals have brought out 0.887 grams.

The ratio

The November 2024 trial retrieved approximately 0.7 grams from Unit 2. A second retrieval from the pedestal floor in April 2025 produced 0.187 grams. Those are the only fuel debris samples recovered, and together they come to under a gram.

Against 880 tons, that is roughly one part in a billion. The figure is not a criticism of TEPCO, because both operations did exactly what they were designed to do: obtain material for analysis so that the properties of the debris can be established before anyone attempts bulk removal. The initial sampling target was under three grams.

880 tonsFuel debris across the three damaged reactors
0.887 gTotal retrieved, across two trial operations
Nov 2024 / Apr 2025The two retrievals, at 0.7 g and 0.187 g
Under 3 gThe initial sampling target, which is the actual goal so far

Radiation, not autonomy, is the binding constraint

News reports of AI robots arriving to help imply the obstacle has been a lack of machine intelligence. It has not been.

Inside the primary containment vessels the dose rates destroy electronics. Semiconductors accumulate damage from ionising radiation: charge trapped in oxide layers shifts transistor thresholds, and single heavy particles corrupt logic or latch circuits into failure. A camera, a controller and a communications link all degrade on a clock measured in hours, and no amount of model capability slows that down.

The equipment reflects this reality. A 72-foot telescoping robotic arm exists to keep its electronics as far from the radiation field as possible. Submersible cleaning robots and hand-sized drones are consumable, sent in with the expectation that some will not come back.

AI can plausibly help with the parts that happen outside the vessel — planning a route, building a three-dimensional model from degraded imagery, deciding where to cut. It does not extend the survival time of the machine that has to go in.

A consortium is an organisational step

The AI robots being reported are, for now, a consortium expected to launch as soon as this fiscal year.

Forming a consortium is a sensible step. Fukushima decommissioning has been criticised for depending on too few contractors, and this is how a government widens the supplier base and brings in robotics groups that would not otherwise bid. It is also the earliest possible point at which this can be announced, and it is several years from a machine entering a reactor.

The distinction between an organisation and a machine matters here. The decommissioning timeline is already measured in decades and has slipped repeatedly. An organisational announcement is not schedule progress, and reporting that treats it as deployment compresses a gap that is the whole difficulty.

What the trial retrievals actually bought

The under-a-gram figure seems absurd, and it reflects the correct sequence of work.

Nobody knows what the debris is. It is a solidified mixture of uranium fuel, zirconium cladding, steel structure and concrete, formed under conditions no one controlled and no one observed. Its hardness, its composition, how it fractures, how much it dusts when cut, and what isotopes come off it are all questions that determine what tooling can be built.

Attempting bulk removal before answering them would mean designing equipment for a material whose properties are guesses. The grams are not the project. They are the specification.

The honest timeline

Nothing in the current sequence supports an expectation of meaningful debris removal this decade.

The steps still ahead are analysis of the samples, design of retrieval tooling against what the analysis shows, qualification of that tooling, construction of containment and handling facilities for material that does not yet have a disposal route, and then removal at a rate that has never been demonstrated. Each has failed to hold a schedule before.

This is not pessimism about the technology. It is the difference between a decommissioning programme and a research programme. Right now, Fukushima is the second one. The consortium may accelerate it. Nothing about a consortium makes 880 tons move.

Why this matters beyond Japan

Several countries in this region are building or expanding nuclear capacity, and the decommissioning liability is the part that gets discounted.

Vietnam has revived its nuclear programme, Indonesia and the Philippines have both signalled intent, and Singapore has commissioned studies on advanced reactors. Every one of those decisions carries an end-of-life cost that is normally modelled from routine decommissioning of an intact plant, which is a well-understood and expensive process.

Fukushima demonstrates what the tail risk of an accident looks like. Fifteen years in, under a gram of the 880 tons has been removed, and a robotics consortium is just now being formed to work out how to proceed. This is not an argument against nuclear power, which has a strong numerical safety record. It is an argument for pricing the accident case honestly. The evidence for that price is being generated right now, 250 kilometres from Tokyo.