what is concurrent loco-manipulation?

It means moving and handling objects at the same time, as one action rather than two. Mopping a floor is the clear case, because the robot must step and push together. Most humanoid systems split the two, walking first and then working while standing still.

how much better is omega-0 than what came before?

On the same 11 chores, with the same training data and the same scoring, omega-0 finished 81.8 percent of runs. The nine baselines ranged from 8.2 percent to 44.5 percent. That is roughly double the best rival, though every number comes from 10 trials per chore.

what does guessing the future actually do?

The model predicts a compact summary of what the scene will look like shortly, rather than drawing a full video. The paper reports that removing this branch cut success from 79.1 to 64.5 percent, and cut success in an unseen room from 79.5 to 15.0 percent. It appears to help most when the setting changes.

can I download the model or the data?

The paper links a project page but does not state that the code or the omega-HOME data set are released. Treat availability as unknown until the team says otherwise. Training the model used 8 NVIDIA H100 GPUs, so reproducing it would not be cheap either.

One robot model did 11 chores, passing 8 tries in 10

What happened

Carry a full cup of tea across a room. You walk, you hold your arm steady, and you watch the floor, all at the same time. You never stop walking to hold the cup. Your legs, your back and your hand work as one.

Most robots cannot do that. They are built in halves. One part walks. Another part reaches. The robot rolls up to the table, stops, then picks things up. Ask it to mop a floor and the split shows at once, because mopping is walking and pushing together.

A team from Singapore and China trained one model to do both at once. They called it omega-0. On a real humanoid robot, in a real room, it wiped tables, mopped floors, emptied a washing machine and shut a drawer with its knee.

The test

The paper is “omega-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation”. Zhe Li, Zhenzhe Zhang, Yangyang Wei and colleagues wrote it, at the MARS Lab at Nanyang Technological University, Peking University, the Beijing Academy of Artificial Intelligence and HKUST in Guangzhou. It went on arXiv on 6 August 2026.

First they had to teach it. A person wore a VR headset with trackers strapped to both feet and drove the robot by moving. That produced a set they call omega-HOME: 40.3 hours of household work across 24 tasks, with video from the robot’s own eyes, video from a camera in the room, and a full record of every joint.

Then they picked 11 chores to test. Pick an apple and drop it in a basket. Wipe a table. Mop a floor. Take clothes out of a washing machine. Sweep paper balls off a bed, turn round, and tip them in a bin. Get a drink from the fridge. Training used about 200 demonstrations per chore, 2,220 runs in all.

One model had to do all 11. No special model per chore. They ran it 10 times on each chore and did the same for nine other well-known robot models, all trained on the same data and judged the same way.

The scoring is careful. A run counts as a success only if every step finishes. They also score part marks, so a robot that grabs the cloth but misses a stain still gets credit.

The result

omega-0 finished the whole chore 81.8 percent of the time. Using only the robot’s own eyes, without the room camera, it managed 79.1 percent.

The nine other models were far behind. The best of them finished 44.5 percent of runs. A well-known copy-the-human method managed 8.2 percent. So the new model roughly doubled the best score on the same chores, with the same training data.

One trick did most of the work. Alongside choosing its next moves, the model also guesses what the room will look like a moment later. It does not draw the picture. It only makes a small guess about it. Switch that guessing off and success falls from 79.1 to 64.5 percent.

The guess matters even more in a new room. Moved to a layout it had never seen, the full model finished 79.5 percent of runs. With the guessing switched off, the same model finished 15.0 percent.

What it means

Two ideas here are worth carrying away.

The first is that a robot does not need to picture the future in detail. Other groups make the robot generate a video of what happens next, which is slow and often wrong. This team keeps a rough sense of what is coming instead. LOOKING AHEAD ROUGHLY BEAT PICTURING THE FUTURE EXACTLY. The model runs at over 7 times a second on the real robot, fast enough to keep its balance while it works.

The second is that the shape of the training data set the ceiling. Every hour came from a person in a headset, moving their own body. That is slow and costly. The team also fed in public human videos, replayed in a simulator so the robot could copy them, which is a way to get more practice without more people in headsets.

Household robots are not arriving in your kitchen this year. But the old excuse, that a robot cannot walk and work at the same time, just got weaker.

Business ideas from this paper

A demo recording service. Robot teams need hours of a person doing chores while wearing a tracked headset. Set up a room, hire and train operators, and sell recorded hours with clean labels. Who buys it: robot startups and university labs with money but no time. Price to test: 200 dollars per recorded hour. One-week test: record five hours for three labs at no charge and see how many order fifty.
A teleoperation kit. The setup here is a headset, two foot trackers, hand triggers and two depth cameras, all wired together. Sell it as one box that works out of the tin. Who buys it: small robot teams who lose weeks getting the parts to talk to each other. Price to test: 3,500 dollars per kit. One-week test: post the parts list and the setup guide, and count how many people ask to buy the built version.
Independent chore scoring. This paper scores each chore in steps and reports every trial. Most robot videos online do not. Offer buyers a plain report: ten runs, this many finished, here is where it failed. Who buys it: investors and buyers judging a robot company. Price to test: 2,000 dollars per report. One-week test: score three public robot demos for free and see who asks for a private one.

How sure can you be?

The work is careful, and it is still small.

This is a preprint. The paper names no journal or conference. Its conclusion lists no limitations, so you have to read the limits out of the numbers yourself.

Each chore was tried 10 times. That is 110 runs for the whole test. Ten tries is enough to see a difference between 8 percent and 80 percent. It is not enough to tell 79.1 percent apart from 81.8 percent. Treat those two figures as the same result.

All of it happened in one place with one robot. Nobody outside the team has repeated it. The team also scored its own runs, step by step, using its own rules.

The averages also hide the hard cases. In the trial tables, simple pick and place finished all 10 runs perfectly. The chore where the robot held a bin and collected three separate pieces of rubbish finished cleanly in only 4 runs out of 10.

The paper links a project page. It does not say the code or the omega-HOME data are released, so treat those as unknown for now.

What would settle it? Another lab, another robot, another house, and the same 11 chores.

Do this today

When someone shows you a robot video, ask two questions. How many tries were there, and how many finished every step? A clip is not a success rate.

Source: omega-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation, August 2026. arXiv:2608.06375 · 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.