ω‑0 Reaches 81.8%: What the Humanoid Home Test Really Shows

A humanoid robot wipes a table, moves sideways while doing so, maintains balance, and adjusts the contact with the cloth – without breaking the task into separate walking and grasping phases. Exactly this simultaneous full-body coordination is the goal of ω‑0. The new research model achieves a success rate of 81.8 percent in a real test with eleven household tasks. The figure is remarkable, but only meaningful if one takes into account the test setup, data basis, and limitations.

Key takeaways

ω‑0 is not a new robot, but a learning control model for humanoid full-body tasks.
The system combines voice command, camera images, and joint states and generates executable movement sequences for the entire body from them.
In the experimental setup, the strongest variant successfully completed 81.8 percent of the full test runs; the average task progress was 90.3 percent.
It was tested on a Unitree G1 with additional dexterous hands – in eleven defined tasks and ten runs each.
The result is a clear research advance, but not yet proof of a freely deployable household robot.

Why household chores are so difficult for humanoids

For humans, wiping a table seems trivial. From a robotic perspective, however, perception, step planning, balance, arm movement, hand strength, and contact with the surface must continuously align. Similarly demanding are picking up laundry from a deep drum, opening a refrigerator, or carrying a container while collecting objects at different heights.

Many previous systems treat locomotion and manipulation as separate modules: First the robot moves to a position, then it stands as still as possible and moves its arms. This works in structured demonstrations, but breaks down in tasks where walking, bending, supporting, grasping, and pulling occur simultaneously. In research, this connection is called loco-manipulation. She is a key to whetherhumanoid robotin the future, to become truly useful in spaces designed by humans.

What ω‑0 does technically differently

ω‑0 is a so-called latent world-action model. It receives a verbal instruction, the current visual observation, and the robot’s own body state. This includes, among other things, joint positions, hand configuration, and torso orientation. From this information, it predicts compact whole-body action representations, which a downstream low-level controller translates into real movements.

The crucial point: The model does not need to generate complete future video frames first. Instead, it learns how the relevant image features of a scene are likely to change. This prediction takes place in a compact feature space. In parallel, a diffusion model generates the next sequence of executable full-body actions. Simply put, the robot does not only learn ‘Which movement comes next?’ but also ‘What should the situation look like afterward?’ — without wasting computing power on photorealistic future images.

Thus, ω‑0 belongs to the next stage of development ofvision-language-action modelsClassical VLAs translate images and language directly into actions. World-action models add a learned understanding of dynamics: they are supposed to capture the consequences a movement has for the body, object, and environment. This is particularly important for contact-rich household work, because small deviations in pressure, grip angle, or stance can cause a long chain of actions to fail.

How to interpret the 81.8% result

The researchers tested seven model families under the same protocol. Each method was trained or fine-tuned on the real demonstration data and had to perform eleven tasks, each with ten independent trials. A run was only considered successful if all subgoals were completed within the time window. The Omni variant of ω‑0 achieved 81.8 percent complete successes, 36.7 out of a maximum of 41 subtask points, and 90.3 percent continuous task progress.

Model	Complete successes	Task progress
Diffusion Policy	15.5 %	40.6 %
π‑0.5	27.3 %	52.8 %
GR00T‑N1.7	22.7 %	49.8 %
ψ‑0	44.5 %	59.6 %
DiT4DiT	43.6 %	61.0 %
ω‑0 Ego	79.1 %	88.7 %
ω‑0 Omni	81.8 %	90.3 %

So the number is not a general “intelligence quotient” for household robots. It is the average within a clearly defined test field developed by the authors. The stronger Omni version uses multiple perspectives during training, including external RGB and depth images; the Ego version works closer to the later onboard camera perspective. According to the paper, no teleoperation, motion replay, or scripted interventions were used during autonomous execution.

Which eleven tasks were tested?

The tests range from short grasping tasks to longer sequences involving movement in space. These include:

pick up an apple and put it in a basket,
place an apple on a shelf,
Throw clothes from the bed into a basket,
put a towel from the basket into the washing machine,
wipe a table,
to mop the floor,
Collect garbage from different heights into a carried container,
throw an apple into a drawer and close it with your knee,
Sweep trash off the bed, turn around, and throw it into a container,
Take clothes out of the washing machine,
to get a drink from the refrigerator.

Ten of the eleven tasks explicitly require lower body movement. Thus, the benchmark tests not only the dexterity of the hands, but the coordinated use of legs, torso, arms, and grippers. This is precisely what distinguishes it from many table-top benchmarks, where a fixed-mounted robot arm works in front of prepared objects.

ω‑HOME: Why the data are almost as important as the model

The performance improvement is not based solely on the architecture. For ω‑HOME, the team collected 40.3 hours of real household data with 4,827 episodes and 24 task types at 30 frames per second. For each trajectory, voice commands, ego camera, external RGB-D recordings, body movements, joint states, and executable action representations were synchronized. For the eleven later tasks, around 2,220 teleoperation demonstrations were gathered, approximately 200 per task.

A human operator controlled the data collection using a VR headset, controllers, and foot trackers. Movements from public human datasets were also replayed in simulation with the SONIC full-body controller. Physically unsuitable trajectories were filtered out by the system. This intermediate step is important: a human movement may look plausible on video, but it can be impossible for a robot with different kinematics, mass distribution, and joint limits to perform.

For companies, this shows a central trend of thePhysical AI: The competition is not just about bigger models. What matters are high-quality real motion data, reliable teleoperation, simulation, a robust full-body controller, and clean feedback from cameras and joint sensors.

Which hardware was in the lab?

The autonomous experiments ran on the Unitree G1. For manipulation, the team equipped the humanoid with Inspire-DexHands. A ZED Mini camera provided the first-person perspective; an additional ZED depth camera captured external RGB-D data. The G1 is thus the physical platform, ω‑0, however, is the research software. Whoever speaks of an “ω‑0 robot” mixes up the model and the hardware.

This separation is important for the evaluation. Whether the method can be transferred to other humanoids with different size, hand kinematics, computing platform, or drive design has not yet been widely demonstrated. Good results on a G1 are a strong signal for the model idea, but not automatic hardware independence. Also, the quality of theSensor fusion, the latency of the control system and the safety functions of the overall platform determine the later practical success.

What the result does not yet prove

First, this is a current pre-publication and not a long-standing reproduced industrial study. Second, the benchmark, data preparation, and evaluation come from the same research project. External replications on other robots and in independent apartments are still missing. Third, the test comprises 110 trials per model – enough for a clear comparison within the project, but too few to derive failure rates for thousands of hours of household operation.

Fourth, the tasks are defined and trained. A real household is constantly changing: light, furniture, packaging, floor friction, pets, children, and unpredictable human movements create an open world. Although the paper reports promising generalization to held-back objects and scenes, this does not yet imply reliable handling of arbitrary new tasks.

Fifth, a high task completion rate says little about safety, noise, energy consumption, wear, cleanability, data protection, or costs. A household product must not only complete a task, but do so repeatedly, quietly, safely, and with low maintenance – and stop in a controlled manner in case of uncertainty.

Why ω‑0 is still an important step

The decisive progress lies less in a single spectacular movement than in unification. One model controls eleven different tasks and links perception, future representation, and whole-body action. It shows that latent prediction can be a practical middle ground between purely reactive policies and computationally intensive video generation.

Three development lines could emerge for the robotics industry. First, teleoperation and data collection platforms will become strategically more important. Second, manufacturers of humanoids are likely to expand standardized interfaces between high-level models and low-level full-body controllers. Third, the industry needs benchmarks that measure not just short success videos, but complete tasks, partial progress, malfunctions, and many repetitions.

Conclusion: Breakthrough in the laboratory, not yet a household product

81.8 percent complete success in eleven real household tasks is a strong result for humanoid loco-manipulation. ω‑0 shows how a robot can learn movement and manual work as a connected problem. Particularly relevant are the compact future prediction, the unified whole-body control, and the carefully synchronized data basis.

The serious conclusion, however, is not that household robots are now ‘solved.’ What has been shown is a powerful research method on a specific platform and in a controlled task environment. The next proof must come from independent tests, other robots, unknown apartments, and significantly longer operating times. If this transfer succeeds, ω‑0 could indeed be among the building blocks that turn humanoids from impressive demonstrators into reliable helpers.

Original sources and technical background

Li et al.: ω‑0 – A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation, arXiv, Version 2 from August 9, 2026
Unitree: technical product page of the G1
Inspire Robots: Manufacturer of the DexHands used in the experimental setup

Bewerte den Beitrag hier!

[Total: 0 Average: 0]

Nico Nuss

Website

Author Nico Nuss has been working on mobile computing and automation software since 2001. Drawing on his experience and strong interest in future technologies, he focuses on robotics and AI.