A humanoid robot becomes useful only when it can move, grasp, balance, and adapt in the same environment. A research collaboration involving Peking University, the Beijing Academy of Artificial Intelligence, NTU, and HKUST Guangzhou has released Omega-0, a model designed to address that whole-body coordination problem. The Omega-0 paper describes a system that turns language instructions, visual observations, and robot state into concurrent locomotion and manipulation actions.
The release matters because many robot systems are still divided into separate components. One model may plan a route, another may control an arm, and another may recognize an object. Omega-0 is built around the opposite idea: a robot should use one policy to coordinate movement and manipulation while reacting to what it sees. The research team calls this setting concurrent loco-manipulation, meaning that a robot walks and handles an object within the same task rather than treating those actions as isolated stages.
Omega-0 Pairs Visual Foresight With Whole-Body Actions
The researchers call Omega-0 a world-action model. Its input combines a natural-language instruction, visual observations, and proprioceptive information about the robot’s body. Its output is not merely a grasp command or a route. It produces actions for the whole body while using a learned estimate of likely visual outcomes to guide the next step.
The paper identifies the core problem as a coupling between navigation and manipulation. A robot cannot reliably pick up an object if it has approached from a poor position. It cannot complete a household task if it walks to the right room but cannot adjust its body to open a cabinet, carry an item, or avoid an obstacle. Omega-0 is intended to connect these requirements through a single policy rather than stitching together independently trained systems.
The model uses what the authors call latent visual foresight. Instead of generating a complete high-resolution future video for every decision, the system predicts a compact latent representation of what a future state may look like. It then uses that representation to help generate whole-body action sequences. This design is meant to give the robot some anticipation of the consequences of a movement without requiring the cost of full video generation at every control step.
The action module is diffusion-based. In simplified terms, it generates a sequence of possible actions and refines it toward a useful motion plan. The paper says this approach helps the model produce coordinated commands for locomotion and manipulation from multimodal observations. The authors position it as a way to handle longer, more realistic tasks rather than narrow demonstrations.
The Omega-HOME Data Set Targets Real Household Tasks
Training data is the other major contribution. The team created Omega-HOME, a data set with 4,827 trajectories collected across 24 task types over 40.3 hours. The tasks are designed to resemble household work rather than a single tightly controlled benchmark. The project page says the evaluations cover 11 real-world tasks and compare Omega-0 with other approaches under the same robotic setting.
That focus on real environments reflects a challenge for embodied AI. A robot can appear competent in a laboratory where object placement, lighting, and walking paths are fixed. It is harder to maintain performance when the robot must move through a room, orient itself around furniture, reach for an object at a different height, and maintain balance while carrying it. The Omega-HOME data set is designed to expose the model to that broader sequence of decisions.
The researchers report an 81.8% real-world success rate in their evaluation. That is a project-reported result, not an independent industry benchmark, but it provides a concrete measure of what the team claims to have achieved. The important detail is the evaluation setting: the results come from physical robot tasks, not only simulated scenes.
The project page also emphasizes that Omega-0 is one unified policy. That matters because every additional module can add a new failure point. A planner may issue a valid instruction that an arm controller cannot execute. A vision model may recognize an object but fail to account for the robot’s balance or location. A unified model does not eliminate those problems, but it tries to reduce the gaps between them.
China’s World-Model Investment Meets Robot Data Constraints
Omega-0 arrives as China’s embodied-AI sector is investing heavily in world models and robot data. EastFrontier reported that China’s world-model startups have raised $5.6 billion for robotics, reflecting investor interest in systems that can help machines reason about physical environments. The research result from PKU, BAAI, NTU, and HKUST Guangzhou shows what that ambition looks like at the model-design level.
The study also exposes why robot data has become strategically valuable. Omega-0 required more than 4,800 physical episodes to train and evaluate its behavior. Collecting that kind of data is expensive because it requires hardware, safe operating setups, human supervision, and task design. China’s robotics companies are increasingly building data-collection facilities and real-world deployment programs partly because embodied models need far more than internet text to improve.
The commercial race is proceeding in parallel. AgiBot’s first-half shipment lead over Unitree shows that domestic vendors are scaling hardware, while programs such as Hangzhou’s robot school are turning structured training environments into an industry resource. Omega-0 is not a commercial product announcement, but it addresses the same practical issue: how to produce behavior that transfers from a training setting into a physical task.
The team’s decision to release a paper and project page gives researchers a concrete technical target to test. The next question is whether the approach generalizes beyond the 24 tasks in Omega-HOME. Real homes, warehouses, hospitals, and service settings create more variation than any one benchmark can capture. Still, the paper’s contribution is clear. It treats robot movement and robot manipulation as one problem, then measures the result on physical tasks.
That is the direction China’s embodied-AI industry needs to move. Building a humanoid body is only the beginning. The larger challenge is teaching that body to predict, adapt, and act coherently in the world around it. Omega-0 offers one research answer: use visual foresight and whole-body generation together, then validate the system where the robot actually has to work.