Jenga Robot: From a Family Game to ABB Arms

Jenga looks like a game for sharp eyes and steady hands, yet good players mostly feel their way through it. They tap a block, sense whether it is loose, and leave it alone if it pushes back. That habit is why the jenga robot became a small landmark in robotics. This post follows the game from a family pastime in Ghana to the boxed game of 1983, then to the MIT robot that learned it on an ABB IRB 120 with force sensing, and on to what touch-based manipulation could mean for robots in factories and, one day, in homes.

A six year old concentrates on pulling a block from a Jenga tower. Photo: Michał Parzuchowski / Wikimedia Commons (CC0)

From Takoradi sawmill blocks to MIT's force-sensing arm

Leslie Scott, a British game designer, played the game with her family in the early 1970s using children's wooden blocks bought from a sawmill in Takoradi, Ghana. She launched it at the London Toy Fair in January 1983 and named it after the Swahili word kujenga, which means "to build". A set has 54 blocks, each 1.5 by 2.5 by 7.5 centimetres, stacked in 18 layers of three. Players pull a block from lower down and put it on top until the tower falls. Pokonobe Associates owns the brand, and Hasbro sells it under licence in most countries.

Friends gather round a table for a game of Jenga in Utrecht, 2014. Photo: Jeroen Komen / Wikimedia Commons (CC BY-SA 2.0)

A camera can spot a block and an arm can reach it, so finding the piece is the easy step for a robot. What the camera cannot see is the friction holding a block in place, the load pressing on it from above, or the moment the tower starts to lean. Human players pick all of that up through their fingertips. Industrial robots spent decades working the other way. They followed exact programmed positions and moved parts that sat where the program expected them. A wobbly stack of wooden blocks, where every pull changes the next one, was a poor fit for that style of automation.

A Jenga tower leaning after many blocks have been pulled, the kind of load a camera alone cannot judge. Photo: Guma89 / Wikimedia Commons (CC BY-SA 3.0)

In January 2019, MIT researchers led by Alberto Rodriguez, with graduate student Nima Fazeli as lead author, described a Jenga robot in Science Robotics. It used a customised ABB IRB 120, one of ABB's smallest six-axis arms, fitted with a soft-pronged gripper, a force-sensing wrist cuff and an external camera. The robot picked a block and a push point, pushed a little, recorded what it saw and felt, and marked each try as a success or a failure. After about 300 attempts it had sorted these tries into clusters of behaviour and could predict which blocks would move. Competing methods needed far more towers to learn the game in simulation.

Students program a small ABB IRB 120, the arm MIT customised for its 2019 Jenga robot. Photo: Peterantala / Wikimedia Commons (CC BY-SA 3.0)

How touch data is training robots in 2026

Rodriguez said at the time that the same skill matters on assembly lines, where the click of a snap-fit or the bite of a screw thread "is coming from force and touch rather than vision." Jenga has since turned into a common test of careful manipulation. At UC Berkeley, Sergey Levine's group built a training method called HIL-SERL, which mixes human corrections with reinforcement learning, and one of its demonstrations had a robot whip a single block out of a tower. The game sits alongside other board game work covered on this site, such as the chess playing robot.

A player eases a block out of a tall, unstable Jenga tower. Photo: Andrew Malone / Wikimedia Commons (CC BY 2.0)

In October 2026 Jenga came back as a way to collect training data. On 20 October, Encord released EMG-1, an open dataset that pairs muscle signals from the wrist with first-person video. Reuters and USA Today covered the work with Jenga as the example: people wear a head camera and wristband sensors while they pull blocks, so the recording tracks hand movement even when the fingers are hidden behind the object. The field set holds 307 hours in 3,804 clips. Encord reports that adding the muscle signals cut median fingertip tracking error by about 25 percent in a lab test. The release is a dataset for researchers, and no robot has played Jenga with it yet.

A Jenga game in progress at home, the kind of hand work Encord records for its robot training data. Photo: Joshjrowe / Wikimedia Commons (CC BY-SA 3.0)

ABB robots already use touch in a narrower, industrial form. ABB's force control lets an arm push or insert a part until it feels a set force, which is how robots fit shafts into bores and grind edges evenly. The GoFa CRB 15000 cobot has torque sensors in its joints, so it stops on contact and can be guided by hand. The dual-arm YuMi IRB 14000 assembles small electronics next to people. Programs like these are usually built and tested first in RobotStudio, and our post on giving ABB robots a sense of touch covers the sensors in more detail.

ABB's Dynamic Assembly Pack uses cameras and force sensing in final car assembly. Photo: ABB

What a touch-trained Jenga robot could do next

Home robots run into Jenga problems all day. A plate wedged under a pan, or a stack of bowls that shifts when one is pulled out, needs a hand that feels before it commits. Robot makers are putting a lot of work into tactile fingers. In October 2026 alone, reports covered a Tesla patent filing for touch-sensitive skin on robot hands, and Shadow Robot and Fraunhofer IFF announced a partnership on tactile-driven manipulation in manufacturing. IDTechEx published an analysis of tactile sensors for humanoid robots the same month, and Interesting Engineering reported on a robotic arm with artificial skin along its limbs.

XELA Robotics uSkin tactile sensors fitted to the fingertips and palm of a robot hand. Photo: XELA Robotics / Wikimedia Commons (CC BY-SA 4.0)

In factories the near-term uses are less playful. The MIT team pointed to separating recyclables and assembling consumer products, and the same feel-then-act loop suits snap-fit assembly and picking parts that lean on each other in a bin. ABB arms could run a learned tactile model, though nobody sells that today. An IRB 120 or a single-arm YuMi IRB 14050 with a wrist force sensor is close to the MIT setup, and SafeMove can limit speed and working zones when people share the cell. For now, any Jenga-playing ABB robot would be a research project built on a standard arm.

A single arm YuMi IRB 14050 in its cell on GE Healthcare's prototype line in Helsinki. Photo: ABB

Tactile sensors are still costly and they wear out, and learned models struggle when a new person, block or surface turns up. Encord itself says generalisation across people, sessions and objects still needs more study. A home robot that pulls on the wrong thing could topple more than a toy tower, so speed and force limits matter as much as clever learning. Training also takes skilled people who collect the data, check it and test the robots safely. Jenga stays a handy yardstick for all of this because a failed move is cheap and everyone at the table can see it happen.

ABB's GoFa cobot is guided by hand and stops when it senses contact. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com