Walking to the job cut a robot’s screw grabs from 87% to 37%

What happened

Think about threading a needle, and sitting at a steady table you manage it. Now try the same thing standing on a moving bus. Same hands, same needle, far worse odds. The job did not get harder, because your platform got wobbly.

That is the problem with humanoid robots at work. Old electric car batteries must be taken apart so the metals inside can be reused, and hundreds of screws come out of each pack. The work is dull and risky, because packs can catch fire, give off fumes or shock a person.

A team built an open system called GOLEM to see whether a humanoid robot could take over, and then they measured what breaks. The answer was not the hands. It was the feet.

The test

The paper is called “GOLEM: Modular Humanoid Autonomy Towards Electric Vehicle Battery Disassembly”. Max Conway, William Xie, Allen Devaraj and colleagues wrote it. They work at the University of Colorado Boulder and the University of Notre Dame, both in the United States. It went online on arXiv on 21 August 2026 as a preprint. No journal or conference is named, but the source code is public on the project page.

The robot is a Unitree H1-2. It stands 1.78 metres tall and weighs 70 kilograms. The battery pack is real, and it is a Hyundai Ioniq 5 pack, 400 kilograms, roughly 1.2 metres by 2 metres. A separate gantry machine loosened each screw first, and the robot then had to pick the loosened screw out of its hole.

The clever part is the staging, and the same grasp was tested three ways. First, hanging in a hoist, so the legs carried no weight. Second, standing on its own feet. Third, standing after walking to the pack by itself. Each step adds one more thing that can go wrong, and every level got 30 tries.

The result

Hanging in the hoist, the robot got the screw 96.7 percent of the time. Standing on its own feet, 86.7 percent. After walking itself into place, 36.7 percent. The hands were fine, and getting the body to the right spot was the problem.

The walking was measured on its own. Over 20 tries at a 6 metre walk, the robot stopped an average of 13.0 centimetres from the target. It reached the goal on 19 of the 20 runs, and that sounds close. However, for pulling a screw out of a hole, 13 centimetres is a mile.

The team also compared two ways of grabbing. One looks once and then reaches, while the other keeps watching through a camera in the palm and corrects as it goes. The look-once method scored 100 percent in simulation and 0 percent on the real pack. The watch-and-correct method scored 96.7 percent in the hoist.

What it means

Here is the lesson worth keeping. WHEN A ROBOT FAILS, ASK WHICH PART FAILED. GOLEM is built so you can swap one piece at a time and watch the number move. That is how the team found the real culprit.

The second lesson is about simulation. In simulation the robot’s body barely moved, by 0.10 millimetres. However, on the real floor it swayed 4.6 millimetres, with a typical peak of 18.0 millimetres during a reach. The authors write plainly that simulation results on precise, contact-heavy tasks do not predict real ones.

The third lesson is about speed. One screw took an average of 75 seconds, from spotting it to holding it, and a person is much faster than that. Therefore, anyone pricing robots against wages today should start from that number.

Business ideas from this paper

1. A docking mat for walking robots. It is a marked floor pad that lets a robot fix its final position before it reaches for anything, so a 13 centimetre stopping error stops mattering. Who buys it: recycling plants and factory integrators running humanoid trials. A price to test: 900 dollars per workstation. A one-week test: send ten robot integrators a one-page before-and-after of stopping error, then count how many ask for a quote.
2. A failure breakdown report for robot pilots. You watch a pilot for a week and split every failure by cause, such as walking, standing, seeing or gripping, the way this paper does. Who buys it: plant managers who are running a robot trial and cannot say why it keeps failing. A price to test: 1,500 dollars per report. A one-week test: do it free for two pilots, publish the charts with permission, then see whether a third site pays.
3. Assembled force-sensing grippers. The paper uses an open-source hand made from printed and off-the-shelf parts for about 460 dollars, and it senses grip force in steps of 0.08 newtons. Most labs do not want to build one. Who buys it: university robotics labs and small integrators. A price to test: 1,200 dollars assembled and tested. A one-week test: list it on one robotics forum, take refundable deposits, and count them.

How sure can you be?

Treat this as an early result, with one robot, in one lab, on one model of battery pack. Each condition got 30 tries. That is enough to see a big gap and not enough to trust a small one.

The robot did not actually unscrew anything, because a separate gantry machine loosened every screw first. The robot only picked the loose screws out, and only two of the five grasping methods were ever run on the real pack.

The authors are careful about this themselves. They write that good numbers in a controlled setting do not show a robot is ready to work near people without supervision. They also declare an interest, because two of them have a stake in a company working on humanoids for manufacturing. The work is funded by an ARPA-E grant.

What would settle it is a whole pack taken apart end to end, in a working recycling plant, timed against a human crew, and repeated over weeks.

Do this today

Therefore, if you are testing any automation, stop counting failures and start sorting them by cause. Ask which single part of the chain the failure belongs to. That one habit is the real finding here.

Source: GOLEM: Modular Humanoid Autonomy Towards Electric Vehicle Battery Disassembly, August 2026. arXiv:2608.21550 · arxiv.org (preprint · not yet peer reviewed).

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.