Industries / Energy, water and waste / Power generation / Nuclear decommissioning
Remote handling
Remote handling is doing work where radiation is too high for people by using tools, arms, vehicles and cameras that operators control from behind shielding or from another building.
6
Milestones achieved, from 1949 to 2021.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1949
Hands behind a wall
Raymond C. Goertz, working for the US Atomic Energy Commission at Argonne National Laboratory, filed a patent on 16 December 1949 for a "remote-control manipulator," a pair of linked arms: an operator moves one and a gripper on the other side of a protective wall copies the motion, so radioactive material can be handled while watched through a shielded window. The patent was granted in March 1953.
Credit
Raymond C. Goertz, Argonne National Laboratory
1984
A rover in the basement
After the 1979 accident at Three Mile Island Unit 2 in Pennsylvania, the reactor building's basement was flooded with radioactive water. William "Red" Whittaker and his students at Carnegie Mellon University built the Remote Reconnaissance Vehicle, which in 1984 became the first vehicle to enter that basement. Fitted with cameras, lights, radiation detectors, vacuums, scrapers, drills and a high-pressure spray, it and a later core sampler surveyed and helped clean the basement for four years, and both remain inside. Carnegie Mellon calls these machines the start of field robotics.
LessonWhittaker kept a second vehicle out of the radiation so it stayed clean for training, tool-building and spares.
Credit
William Whittaker and students, Carnegie Mellon University
1986
Moon rovers on the roof
After the Chernobyl explosion, the Soviet institute VNIItransmash built the STR-1, a remote-controlled bulldozer based on its Lunokhod lunar rover work, in under two months, with its electronics in a lead-shielded box and its circuits built from relays. From 8 August to 10 October 1986 it pushed radioactive debris on the roofs of Unit 3 and the shared ventilation block. The same account says the robots "did not play a determining role" in clearing the roofs, and that the work was done mainly by people, among them soldiers and civilian specialists. A 2026 summary of a documentary series on the clean-up reports that many robots were sent in untested, that one overstated radiation readings by more than ten times and that two tipped over.
Lessonrobots rushed in without testing did little of the work, and their builders concluded that such robots must be more reliable and must be set up, serviced and recovered without people.
Credit
VNIItransmash (Leningrad)
2011
Borrowed robots go first
On 17 and 18 April 2011, five weeks after the Fukushima Daiichi accident, two PackBot robots from the US company iRobot entered the reactor buildings of Units 1, 2 and 3 and sent back the first measurements from inside, including up to 49 millisieverts an hour in Unit 1 and 57 in Unit 3, as reported by IEEE Spectrum. Japan, a country known for robots, relied on machines from abroad because its own had not been tested as extensively, and a TEPCO spokesman said the company had not expected to use robots until iRobot offered them.
Lessonrobots for an accident need to be built, tested and practiced with before the accident.
Credit
iRobot (robots); TEPCO (operator)
2017
Scorpion stalls, sunfish swims
In February 2017 a Toshiba robot shaped like a scorpion, about 54 cm long, entered the containment of Fukushima Daiichi Unit 2, but its crawler failed on radioactive debris and it was left inside where it would not block the next probe. Earlier that month a cleaning robot sent ahead of it lost two cameras after about two hours, at the 1,000-sievert total dose they were built to take, because radiation was higher than expected. In July 2017 a 30 cm swimming robot nicknamed Little Sunfish, built by Toshiba and IRID to fit a 14 cm opening and tolerate about 200 sieverts, entered the flooded containment of Unit 3 and sent back pictures of damage not seen before.
Lessona robot for a damaged reactor needs margin for conditions far worse than estimated, and a plan for where it can be left if it fails.
khaosodenglish.comworld-nuclear-news.orglearningenglish.voanews.com
Credit
Toshiba and the International Research Institute for Nuclear Decommissioning (IRID); TEPCO (operator)
2021
A robot dog at Sellafield
In October 2021 the Sellafield nuclear site in England brought in Spot, a four-legged robot made by Boston Dynamics, and by December it had completed its first shift in a contaminated area, collecting laser scans and gamma radiation images while wearing a protective suit. By November 2023 Sellafield was using Spot routinely for inspections and for sorting wastes, said it was the first in the world to use a laser scanner (lidar) on a remotely operated vehicle in a high-radiation area, and reported the same robots in use at other NDA sites, including Dounreay.
Credit
Sellafield Ltd, with Createc, the UK Atomic Energy Authority and Boston Dynamics
Watch for
A UK arm for Fukushima
IRID says it has finished developing an 18-joint robot arm, 22 m long and weighing 4.6 tons, based on arm technology from Veolia Nuclear Solutions in the UK that was first used for fusion reactor maintenance. The arm was built in the UK from autumn 2018, shipped to Japan in July 2021, tested on a mock-up at the Japan Atomic Energy Agency's Naraha center, and moved to Fukushima Daiichi on 7 April 2026. It is meant to open a route into the Unit 2 containment, scan the inside with a 3D laser scanner and take out fuel debris on a trial basis.
Achieved whenTEPCO reports that the arm has reached the area below the reactor vessel and brought out a sample, and Japan's Nuclear Regulation Authority or an outside reporter confirms it.
Claimed by
IRID, April 2026
Robots inside gloveboxes
A glovebox is a sealed box with built-in gloves for handling radioactive material; Sellafield has hundreds that are no longer needed. In July 2025 Sellafield and AtkinsRéalis said they had fitted a robot arm, the ARGO system, inside two working gloveboxes, with engineers operating it, and called this a world first. In May 2026 Sellafield said it had taken a glovebox from installation through full remote size reduction and packing with robots and laser cutting, another world first by its own account, with a second glovebox under way.
Achieved whenrobotic dismantling is used on many gloveboxes, and the regulator or an independent review reports the number done and the dose saved.
Claimed by
Sellafield Ltd and AtkinsRéalis, July 2025 and May 2026
Open
01
Learned steps in a glovebox
A non-radioactive glovebox mock-up with a robot arm inside, and policies learned from teleoperated demonstrations for three dismantling steps: undoing bolts on a panel, cutting a plastic tube with a powered shear, and placing the pieces in a waste drum mock-up. Publish the tally of attempts for each step.
The briefCredit · unclaimed02
Public labels for containment video
An open set of labeled frames from the videos TEPCO has published of its robot investigations inside the damaged containments at Fukushima Daiichi, with classes such as structures, deposits and corrosion, a held-out test split and baseline scores, so any team can test perception models on real post-accident footage.
The briefCredit · unclaimed
Open 01 · the brief
Learned steps in a glovebox
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A non-radioactive glovebox mock-up with a robot arm inside, and policies learned from teleoperated demonstrations for three dismantling steps: undoing bolts on a panel, cutting a plastic tube with a powered shear, and placing the pieces in a waste drum mock-up. Publish the tally of attempts for each step.
Why it is still open
At Sellafield, engineers operated the ARGO arm. The UK Atomic Energy Authority published an open simulator of a robot arm in a glovebox in 2020, for work on remote control and autonomy, which runs only in software. A University of York project trained an arm by reinforcement learning to pick and place 3D-printed "fuel pellets" in a mock glovebox, and a 2025 paper learned a plug-insertion task from teleoperation for nuclear waste handling. No published tally for dismantling steps learned from demonstrations was found.
Built with what exists
an acrylic glovebox mock-up; a UR5e or a pair of low-cost SO-101 arms; LeRobot's open training and data tools (see the Note "No download"); open robot models such as Runway's Praxis-1 (see the Note "Open robot model weights") or NVIDIA's GR00T 2 (see the Note "Five demonstrations per task").
First step
Build the mock-up around one arm and record 50 demonstrations of undoing four bolts.
Done looks like
For each of the three steps, 20 filmed attempts by the learned policy with a published success tally and the number of demonstrations used; data, code and trained weights released.
Open 02 · the brief
Public labels for containment video
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open set of labeled frames from the videos TEPCO has published of its robot investigations inside the damaged containments at Fukushima Daiichi, with classes such as structures, deposits and corrosion, a held-out test split and baseline scores, so any team can test perception models on real post-accident footage.
Why it is still open
The UK Atomic Energy Authority's LongOps program labeled past containment videos into 13 classes and trained a detection model, but its 2024 paper describes no public release. Researchers have built radiation-noise image sets and noise simulators, which address camera noise rather than what the robot is looking at.
Built with what exists
TEPCO's published investigation videos, used under its terms; CVAT or Label Studio for labeling, with Meta's SAM 2 to speed outlining; a small vision model fine-tuned on one graphics card (see the Note "One GPU to fine-tune"); the University of Tokyo's open gamma-noise simulator (github.com/matsuren/pixel_noise_sim_geant4) to test models against heavier noise.
First step
List every containment investigation video TEPCO has published, with date, unit and robot, and write to TEPCO asking whether labeled frames may be shared.
Done looks like
At least 3,000 labeled frames from at least ten videos; a second person relabels 300 frames and agrees on at least 90%; baseline scores published; the set released under terms TEPCO accepts.