A research team spanning Nanyang Technological University, Peking University, HKUST, and BAAI has built a single AI model that lets humanoid robots walk, see, and manipulate objects at the same time. The model, called Omega-0, achieved an 81.8 per cent success rate across 11 real-world household tasks — outperforming every competing system including pi-0.5, EgoVLA, GR00T-N1.7, and psi-0.
The paper, published on arXiv on August 6, tackles a problem that has limited humanoid robotics for years: most AI models treat walking and grasping as separate tasks. A robot walks to a shelf, stops, then reaches. Omega-0 does both simultaneously — the kind of coordinated movement humans do without thinking.
Why concurrent loco-manipulation matters
What is loco-manipulation? It’s the ability to move your body while using your hands. When you walk across a kitchen and grab a cup off a shelf in one fluid motion, you’re doing concurrent loco-manipulation. Your legs adjust your position, your torso leans, your arm reaches — all at once, coordinated by a single brain.
Existing humanoid robots can’t do this well. Their AI policies decompose movement into stages: walk to location, stop, plan grasp, execute. It works for structured factory tasks where the robot moves to a fixed position and picks up a known object. It fails in homes, where you need to reach for something on a high shelf while stepping around a chair.
The arXiv paper frames this as the core gap between industrial humanoids and household humanoids. Factory environments are predictable. Homes are not. A robot that can walk-and-grab simultaneously handles clutter, unexpected obstacles, and tasks that require adjusting posture mid-movement.
How Omega-0 works
Omega-0 is a latent predictive world-action model. Instead of predicting future video frames (the approach taken by many recent world models), it learns compact embeddings of future observations — a lighter computational target that still captures what the robot needs to know about what happens next.
The model takes three inputs: a language instruction (“put the cup on the table”), current visual observation (from egocentric or exocentric cameras), and the robot’s proprioceptive state (joint positions, velocities). It outputs whole-body action latents that are directly compatible with the robot’s controller — no separate translation step.
The team also collected Omega-HOME, a 40+ hour dataset of real-world household humanoid tasks with synchronized multi-view observations, whole-body SMPL motion captures, robot states, and action latents. This dataset grounds the model in real physical data rather than simulation-only training.
The 81.8 per cent number in context
The 81.8 per cent success rate sounds high, but the context matters. The 11 tasks were real household chores — not simulation tests. The competing systems all scored lower on the same tasks. And the single-model approach means one Omega-0 handles all 11 tasks, not 11 task-specific models.
What the paper doesn’t claim is generality beyond these tasks. Eleven household tasks is a meaningful benchmark, but it’s not “any household task.” The tasks likely include structured variations of reaching, carrying, placing, and opening — the building blocks of home work, not the full range.
The comparison against pi-0.5 and GR00T-N1.7 is notable because those are production-grade systems from well-funded labs. Omega-0 beating them on real hardware — not simulation — suggests the concurrent loco-manipulation approach has genuine advantages over the decomposed pipeline.
What this means for the humanoid industry
The humanoid robotics industry has split into two camps. The deployment camp — Agility Robotics, Tesla, Figure AI — is building robots that work in factories and warehouses now, with narrow task sets. The research camp — NTU, Peking University, DeepMind, NVIDIA GR00T — is building the AI models that will eventually let those robots work in homes.
Omega-0 sits in the research camp, but it addresses the specific bottleneck the deployment camp faces: robots that can only do one thing at a time are slow. A warehouse robot that walks, stops, picks, walks, stops, places is half as productive as one that walks-and-picks in a single motion. The concurrent approach isn’t just a home-robotics breakthrough — it’s a throughput improvement for industrial deployments too.
The China-Singapore research collaboration
The research team spans three countries: Nanyang Technological University (Singapore), Peking University and BAAI (China), and HKUST (Hong Kong). This is a cross-border academic collaboration, not a single-lab effort. The Omega-HOME dataset — 40 hours of real humanoid household data — is itself a contribution, as most humanoid datasets are either simulation-only or proprietary.
The paper was submitted August 6 and revised August 9, suggesting active iteration. It’s available on arXiv as 2608.06375, and the authors have released the dataset name (Omega-HOME) and methodology details.
❓ FAQ
What is Omega-0? A latent predictive world-action model that lets humanoid robots perform locomotion and manipulation simultaneously, rather than in separate stages. It was developed by researchers from NTU, Peking University, HKUST, and BAAI.
What does 81.8 per cent success rate mean? Across 11 real-world household tasks performed on physical hardware, Omega-0 completed 81.8 per cent of attempts successfully. Competing models (pi-0.5, EgoVLA, GR00T-N1.7, psi-0) all scored lower on the same tasks.
When will this be in commercial robots? Not soon. Omega-0 is a research model, not a product. The contribution is the architecture and dataset, which other labs and companies can build on. Commercial deployment would require integration with a specific robot platform, safety certification, and scaling to a wider task range.
How is this different from what Tesla or Figure AI are doing? Tesla and Figure AI are deploying robots in factories with narrow task sets. Omega-0 is a research model focused on the AI problem of concurrent movement and manipulation — the capability gap between factory robots and home robots.
🔍 THE BOTTOM LINE
Omega-0 is a research breakthrough, not a product. But it addresses the right problem. The gap between a robot that works in a factory and one that works in a home is not strength or speed — it’s the ability to move and act simultaneously in unstructured spaces. An 81.8 per cent success rate on 11 real tasks, beating production-grade competitors, is a meaningful step. The question is whether the deployment-focused companies pick up this architecture and run with it.
