en.Wedoany.com Reported - Chinese embodied intelligence company Wujie Dongli officially released the MWA™ Embodied General Brain on June 29. The model is positioned as a "long-sequence bidirectional physical causal chain" latent space world model. In the RoboCasa GR1 TableTop embodied intelligence benchmark, the MWA™-WALA, jointly released by Wujie Dongli and the Institute of Automation, Chinese Academy of Sciences, achieved a global first-place average task success rate of 75.2%, surpassing models such as NVIDIA GR00T-N1.6, Daxiao Robot ACE-EGO-0, XPeng DIAL, and AutoBot ABot-M0. This release further advances Wujie Dongli's previously emphasized "latent space world model + reinforcement learning" approach towards long-cycle, multi-step, high-precision manipulation tasks.
The role of a world model is to enable a robot to understand environmental changes and the consequences of its actions before executing them. For embodied intelligence, the model must not only "see" objects but also determine how they will move, collide, slide, bear force, and stabilize.
The core change in MWA™ lies in modeling physical causal relationships within the latent space. Traditional robot models often rely on direct mappings between vision, language, and action. While they can perform short tasks like grasping, placing, and switching, they are prone to issues in long-cycle operations such as action sequence breaks, path drift, unstable force application, and misjudgment of target states. MWA™ employs a temporal chunk-level inverse dynamics modeling mechanism, outputting continuous multi-step Latent Action Chunks. This allows the robot to generate not just single-step actions but a continuous action chain centered around a task process. Consequently, when performing multi-step operations, the robot can process the "current action," "next state," and "final task goal" within the same temporal framework.
The AnyPhys negative sample core data system is designed to supplement the robot's understanding of failure boundaries. Embodied intelligence training cannot rely solely on successful demonstrations; real-world operations more commonly involve slipping, collisions, mis-grasps, posture instability, object spillage, and misoperations in confined spaces. MWA™ uses a large number of failure, instability, and critical boundary samples to inversely deduce the physical stability domain under different working conditions, enabling the robot to avoid high-risk paths when generating actions. For industrial and commercial service scenarios, this capability is more critical than simply improving demonstration performance, as robots need to maintain stable, safe, and reproducible operation over extended periods.
The RoboCasa GR1 TableTop primarily tests a robot's generalization ability in tabletop tasks. Complex objects, non-standard scenes, lighting variations, and clutter interference amplify differences in the model's spatial understanding, action continuity, and task planning capabilities.
The 75.2% average task success rate indicates that MWA™ has achieved high performance in tasks such as multi-step coherent operations, object retrieval in confined spaces, and precise picking of scattered items. The value of simulation environments like RoboCasa lies in their ability to test robot policies across a large number of kitchen, tabletop, and object interaction tasks, rather than just evaluating a single grasping action. MWA™'s top ranking signifies that the latent space world model approach is highly competitive in simulation benchmarks and provides a new algorithmic validation foundation for subsequent real-world robot deployment. Official RoboCasa documentation states that its benchmark is used for comparing generalist robot policies, with tasks covering multiple types of operations in everyday environments and focusing on multi-task learning and generalization evaluation.
This release also continues Wujie Dongli's recent technological and commercial momentum. Wujie Dongli had previously completed over $200 million in angel round financing and built the embodied general brain MWA™ around the "latent space world model + reinforcement learning" approach. The company's second-generation robot, K15, has entered batch production, with application directions covering industrial manufacturing, commercial services, and other scenarios. For Wujie Dongli, the model's benchmark ranking is merely a technological entry point. The real test lies in whether MWA™ can enter real factories, commercial spaces, and long-cycle mobile manipulation tasks, withstanding the engineering pressures of production line rhythms, open environments, safety constraints, and multi-task switching.
Embodied intelligence is moving from "being able to demonstrate" to "being able to operate stably over the long term." If MWA™ can transfer its long-sequence action capabilities from simulation to real robot hardware, it will help improve robot reliability in assembly, sorting, pick-and-place, service, inspection, and complex spatial operations. For the entire industry, competition among such models is no longer just about language understanding or visual recognition capabilities, but about physical world modeling, action continuity, failure sample learning, and closed-loop iteration in real-world scenarios. With this release of MWA™, Wujie Dongli has also established the latent space world model as an important technical branch in the path towards an embodied general brain.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
