Embodied AI shifts from data hoarding to world models as robots master long-horizon tasks

According to a report by QbitAI, Chinese startup Moqi Intelligent demonstrated a wheeled robot named MORPHI KINO at WRC 2026 performing a continuous 15-minute household routine. The robot cleaned a table, checked the refrigerator inventory and restocked a bottle of water, moved a wet shirt from a washing machine to a dryer, and then folded a dry T-shirt—all without human teleoperation. Moqi attributed the performance to its self-developed model architecture MoRA, which makes the robot “agentic-native.” Unlike conventional systems that rely on a large language model to plan every sub-step, MoRA embeds goal persistence, multi-level memory, and progress awareness directly into the low-level policy model, allowing the robot to recover from deviations without frequently consulting a higher-level planner.

At the same event, UBTech showcased multiple humanoid robots across industrial, commercial, and home scenarios, according to a press release published by QbitAI. The industrial robots Cruzr Y1 and Cruzr S2 performed tasks such as loading and unloading automotive parts and sorting parcels. The company also unveiled a full-size commercial service robot Walker C1 and a super-realistic humanoid robot U1 equipped with an emotion model. UBTech said its in-house foundation model Thinker and world model Thinker-WM underpin these products, with the latter topping the Libero benchmark for embodied intelligence.

On the data infrastructure side, 51WORLD launched AperData, a hardware-software system for embodied data production, at an August 18 event. According to the release, the first-generation product is priced at 5,100 yuan per set and allows human workers to wear an AperEgo capture device to record first-person multimodal data, significantly lowering the cost and complexity of robot data collection. The accompanying AperOS platform performs on-device quality checking and cleaning, discarding invalid clips before uploading, which the company claims improves data production efficiency tenfold compared with traditional teleoperation.

Academic researchers are also rethinking data needs. At RSS 2026, Stanford PhD student Wenlong Huang (working with Fei-Fei Li) described a new approach called Point-World, according to Leiphone. He argued that robots should not rely on memorizing millions of demonstration videos. Instead, they can use a 3D point-cloud world model to simulate counterfactual outcomes—“what would happen if I move this way”—and search through possible actions before acting. Huang showed that with only 15 hours of 3D interaction data fine-tuned on top of 2D video pretraining, the model achieved better dynamics prediction than a model trained on 500 hours of 3D data. This points toward recursive self-improvement, where robots continuously acquire new behaviors without human demonstrations.

Nvidia also unveiled its Cosmos 3 world foundation model at an RSS workshop, as reported by Leiphone. Cosmos 3 is described as the first omni-modal model to unify text, vision, audio, and action in a single Transformer architecture. It can serve as a world simulator, an inverse-dynamics model, and a policy model. The company released open-source versions, including an edge model (Cosmos Edge, 4B parameters) that runs in real time on Jetson devices, and larger models up to 64B parameters.

These developments across conferences and companies indicate that the embodied AI field is shifting from an emphasis on massive datasets to mechanisms that let robots imagine and generalize. Data-quality-focused tools like AperData and model-level innovations such as Point-World and MoRA are trying to solve the same bottleneck: how to make physical-world learning efficient enough for deployment.