After Raising Hundreds of Millions in Funding, Top Rankings Achieved: WuJie Dynamics Releases MWA Latent Space World Model

On June 29, Wujie Dynamics officially released the MWA Embodied General Brain, a latent space world model featuring "long-horizon bidirectional physical causal chains." Simultaneously, it set a new industry record and ranked first globally in the RoboCasa GR1 TableTop benchmark for embodied AI, jointly initiated by Stanford University and other institutions.

MWA eliminates the redundant noise and high computational costs associated with pixel-space prediction, performing all inference within a unified shared latent space.

Furthermore, MWA implements a temporal chunk-level inverse dynamics modeling mechanism, breaking through the limitations of traditional world models that rely on "single-step latent action inference." It reconstructs the output paradigm of inverse dynamics models, endowing them with long-horizon causal induction capabilities to batch-infer and output continuous multi-step Latent Action Chunk groups.

Currently, industry datasets generally suffer from common issues such as "prioritizing positive over negative samples" and "sample homogeneity," with the vast majority consisting purely of positive samples or containing only a minimal amount of negative samples. This singular sample structure cannot support the dense reward training required by reinforcement learning. Without multi-dimensional sample comparisons and boundary constraints, models are highly prone to decision paralysis and insufficient generalization when facing abnormal disturbances in real-world operating conditions due to cognitive gaps.

Addressing this issue, Wujie Dynamics pioneered the AnyPhys negative-sample core data system. It deeply intertwines deep negative samples, fine-grained boundary instability samples, suboptimal samples used for policy alignment, and baseline positive samples to construct a physically bounded coordinate system with high information density, thereby filling the gap in full-dimensional sample shortages required for dense reinforcement learning training.

Previously, on June 26, Wujie Dynamics announced the completion of an angel-round financing exceeding $200 million.

After raising hundreds of millions of dollars in funding, the company has topped industry rankings and unveiled the MWA Hidden Space World Model. The latest angel round was jointly invested by JD-affiliated funds, C Capital, Hongyi Investment, Shengyu Investment, Fengyuan Investment, and other institutions, with follow-on investments from existing shareholders including Linear Capital, Sequoia China, Huaye Tiancheng, and Yairui Capital. The raised funds will continue to be used for the R&D of a general-purpose embodied AI brain, the construction of technological infrastructure, and global large-scale delivery. Meanwhile, the Pre-A round of nearly $200 million in financing is also nearing completion.

In terms of commercialization, Wujie Dynamics has signed global orders totaling nearly $100 million, establishing deep cooperation with Envision Group, ZF LIFETEC, Omron Group, and well-known domestic and international coffee chains, among others. This collaboration aims to advance the global deployment and continuous evolution and iteration of general-purpose embodied AI across diverse real-world scenarios.

Furthermore, with the second-generation robot K15 already entering mass production, Wujie Dynamics is about to fully enter its global delivery cycle.