Omega-0: The Model Teaching Humanoid Robots to Walk and Manipulate Objects Together

Omega-0 (ω-0) is a unified whole-body world-action model for humanoid robots. Rather than treating walking and object manipulation as separate challenges, it attempts to coordinate them within a single policy. A robot can move its base, adjust its posture and viewpoint, maintain balance and use its hands as part of the same task. 1

The model’s central idea is to predict future visual features instead of directly producing one long, predetermined trajectory. Put simply, Omega-0 estimates what the robot is likely to see after taking an action, then uses that prediction to select movements that work together. The reported results come from the researchers’ paper and should be understood as the authors’ claims, not as independently verified benchmark results. 1

How Omega-0’s three-stage architecture works

1. Compressing whole-body movement and learning representations

The first stage converts high-dimensional whole-body movements into more compact action tokens, or latent representations, using a Whole-Body FAST tokenizer. This gives the model a more manageable way to represent complex combinations of leg, torso and arm movements. 1

The researchers also adapt a Qwen3-VL-2B-Instruct vision-language backbone to the robot setting. Training uses learnable tokens that distinguish between the robot’s egocentric, or first-person, view and an exocentric, or third-person, view from an external camera. This allows the model to benefit from a wider third-person perspective during training while relying on the robot’s own camera during operation. 1

2. Predicting future visual states before acting

In the second stage, Omega-0 uses a V-JEPA-inspired latent-prediction objective. It combines visual information, a language instruction and proprioception—the robot’s measurements of its own physical state—to predict future visual features rather than reconstructing every future pixel. 1

A diffusion transformer then uses those predicted representations to generate a sequence of whole-body action latents that the robot’s controller can execute. The intended effect is that an action is chosen not merely because it is mechanically possible, but because it is expected to lead to a useful visual state. For example, the robot may move closer, orient its body and camera, maintain its balance and then reach for an object. 1

3. Grounding the model in robots and real homes

In the third stage, public human demonstrations are transferred into the robot’s action-latent space through simulation-based grounding. The system is then adapted using real data from humanoid robots performing household tasks. 1

During execution, Omega-0 does not commit to the entire trajectory of a long task. Instead, it generates a short action chunk, executes it, observes the environment again and replans. This receding-horizon approach allows the robot to correct its course using new visual and proprioceptive information rather than following a completely open-loop trajectory. 1

What the Omega-HOME dataset contains

The paper introduces Omega-HOME, a dataset of real household interactions collected with humanoid robots. It contains 40.3 hours of data, 4,827 tele-operated episodes and 24 household tasks, recorded at 30 hertz. 1

The synchronized data include:

• language instructions;
• egocentric RGB video from the robot’s camera;
• exocentric RGB-D video from a ZED camera;
• proprioceptive robot states;
• whole-body motion references; and
• controller-compatible action latents. 1

The tasks are organized into eight capability groups:

• object retrieval;
• surface cleaning;
• appliance interaction;
• transferring objects between containers;
• cloth handling;
• arranging items in storage spaces;
• mobile manipulation; and
• using tools for floor-based tasks. 1

The dataset therefore goes beyond static tabletop grasping. Many activities require the robot to move its base, adjust its posture and torso, maintain balance and manipulate furniture, objects or tools over longer sequences. 1

How the researchers addressed possible benchmark contamination

Under the protocol described in the paper, the 11 task types used for evaluation were removed from the Omega-HOME training data. The remaining robot data were combined with public human demonstrations for pretraining, while real household data were used for grounding and post-training. 1

This reduces the risk that the model simply memorized the same demonstrations later used in testing. It is not, however, an absolute guarantee of generalization. Training and test environments may still share similarities in rooms, objects, task structures, robot hardware or data-collection procedures. The stronger test would be independent replication in new homes and on different humanoid platforms.

Reported performance on real household tasks

Across 11 real-home loco-manipulation tasks, the authors report an 81.8% average success rate for Omega-0. The result uses one unified model rather than separate policies for each task, different action heads or independent modules for walking and manipulation. 1

The evaluations covered activities including tabletop manipulation, cleaning, laundry, object transfer, appliance interaction and mobile manipulation. The paper also reports that Omega-0 outperformed the compared baselines π-0.5, EgoVLA, GR00T-N1.7 and ψ-0 on this test suite. 1

The available material does not provide sufficiently clear evidence for the exact overall percentage achieved by each baseline. The defensible conclusion is therefore the reported ranking and Omega-0’s 81.8% result—not a more detailed numerical comparison that is not documented here.

What 81.8% means for embodied AI

Omega-0’s main contribution is architectural. It points toward a direction in which predicting future perceptual latents can connect human demonstrations, language, vision, proprioception and whole-body robot control. 1

Instead of learning only which movement corresponds to a particular image, the model attempts to connect an action with the state it is expected to produce. That is especially useful for tasks in which a robot must decide at the same time where to stand, what to look at and how to use its hands.

The result does not mean humanoid robots are ready for unsupervised, general-purpose operation in homes. Several major challenges remain:

• Long-horizon reliability: Success across 11 tasks does not demonstrate dependable operation for many hours, recovery from accumulated errors or robust execution of chains of interdependent subgoals.
• Safety: Interaction with people, fragile objects, stairs, clutter, pets and unexpected movements requires explicit safeguards, uncertainty management and hardware-level verification.
• Generalization: Performance on held-out objects and scenes is encouraging, but it does not settle whether the system will transfer reliably across different homes, lighting conditions, layouts, object-placement habits or robot morphologies. 1
• Adaptation: A real home is constantly changing. A practical system must detect failures, ask for help when necessary, learn safely and adapt without expensive retraining.

The most accurate takeaway is that Omega-0 represents a significant step toward unified locomotion and manipulation in humanoid robots. Its reported 81.8% is a strong research result on a specific benchmark—not proof that a robot can already perform every household task reliably without supervision.