Greater Bay Area’s First Embodied AI Unicorn Breaks 200 B RMB Valuation

World‑Unified‑Model (WUM) and WALL‑B

Self‑Variable released WALL‑B in April, the first embodied large model built on a World‑Unified‑Model (WUM) architecture. The model jointly trains vision, language, action and physics prediction in a single network, removing inter‑module loss. Synchronous annotation of multimodal data (visual, auditory, language, tactile, motion) enables “multimodal‑in, multimodal‑out” inference without translation between modules.

WALL‑B provides native body awareness: the model can infer its own spatial dimensions (height, arm reach, joint limits) without external sensors, allowing self‑assessment of passability through narrow spaces and reachability of objects.

WALL‑B also incorporates a native physical “world view”, delivering strong zero‑shot generalisation and a closed‑loop self‑evolution capability.

Open‑source releases

WALL‑OSS‑0.5 – an open‑source model pretrained on the same data. In evaluation on 17 real‑robot tasks, it achieved >80 % success on 4 tasks without any fine‑tuning, outperforming mainstream open‑source models on operation and inference tasks. It demonstrates that a pretrained state can be directly deployed on hardware with performance comparable to fine‑tuned models.

WALL‑WM world model – the first world model with event‑level prediction. It replaces uniform time sampling with event‑aligned multimodal data, enabling the model to understand physical evolution at the event granularity.

XRZero‑G0 data pipeline – an end‑to‑end solution for body‑free data collection, automatic quality inspection, mixed‑training data generation and real‑device evaluation. Experiments showed a 95 % reduction in effective data acquisition cost (to 1/20 of traditional real‑robot collection).

Full‑stack technology matrix

The stack consists of a closed‑source flagship, an open‑source base, and an open‑source world model, forming a “model‑data‑embodiment” closed loop.

Hardware deployment

On May 25 a new wheeled dual‑arm home robot equipped with WALL‑B entered the first batch of households. The robot supports object sorting, basic cleaning and item delivery. On‑device visual de‑identification ensures raw images never leave the device, satisfying privacy requirements.

Collaboration with 58.com launched a human‑robot collaborative cleaning service in Shenzhen and Beijing. Robots handle living‑room sorting and basic cleaning; humans perform communication and deep cleaning. Service pricing matches pure‑human cleaning, providing real‑world interaction data for continual model improvement.

The technology has also been validated on automotive parts production lines, demonstrating applicability in industrial scenarios.

Remaining challenges

General‑purpose embodied intelligence remains in early large‑scale household adoption. Open questions include long‑term user feedback, commercial replication efficiency, open‑source ecosystem activity, cost, reliability and overall user experience.