Anu Intelligence Releases Six Embodied AI Models in One Week to Close Industrial Deployment Loop

QbitAI reported that China's embodied AI sector had raised about 43.8 billion yuan by June 12, 2026, nearly 80 percent of the full-year total for 2025. The report said many demos and financing rounds cover home companionship, commercial guidance and industrial production lines, but running a demo and deploying at scale are different. It argued industrial settings are closest to batch deployment because tasks such as picking, placing, moving and plugging are repetitive and have relatively fixed boundaries.

Data from the State Taxation Administration showed that in the first five months of 2026, Chinese industrial enterprises increased their purchases of embodied AI robots by 2.3 times year on year, according to QbitAI. The report said spending is shifting from buying demos to putting systems on production lines. Traditional industrial robots already handle welding, transport and painting, but they face hardware limits because fixed structures must be redesigned for new workstations or materials, and software limits because they rely on pre-programming and cannot understand scenes, instructions or generalize.

The six models are Helios, SimLab, ActiWorld, EvoHIL, AnuVerse-0.5 and Nexus. Helios is a real-time long-video generation model that can generate minute-long videos and run inference at 19.5 FPS on a single H100 GPU, QbitAI reported. Although Helios was not originally designed for robotics, it addresses long-horizon modeling, continuous generation, real-time inference and efficient video generation, and has been adopted by several leading embodied AI and world-model teams.

SimLab is a soft-object simulation platform. It focuses on grasping-data generation for logistics soft-package sorting, thermal-vision data generation for industrial equipment and sim-to-real transfer. QbitAI reported that the platform has been accepted in actual projects and that a full workflow has been run with a hundred-yuan-class robotic arm and zero real-world annotation. Its core includes a self-developed soft-body and thermal physics solver, a synthetic data factory built on Isaac Sim, a sim-to-real transfer method validated on real robotic arms and a self-developed data-collection kit.

ActiWorld is a world model designed to make robots accountable for the consequences of their actions before they move. It considers action, time and space, and uses three training constraints: an inverse dynamics head, bidirectional prediction and motion weighting. These constraints are used only in training; inference uses standard interfaces and remains compatible with existing action-conditioned diffusion backbones, QbitAI reported. In tests, ActiWorld's judgment of success or failure differed from real outcomes by less than 3.25 percent. In a sorting task that required a package barcode to be placed face-up on a conveyor, the original policy had a 56 percent success rate; with ActiWorld-guided rehearsal, the relative failure rate dropped by about 13.6 percent, according to the report.

RoHIL and EvoHIL are two generations of real-machine reinforcement learning technology. Under these frameworks, QbitAI reported, robots can reach full-score performance in 30 minutes of training and do not need new data collection when lighting changes. Reward criteria no longer require manual offline relabeling each time; the system iterates from human-confirmed successful samples without stopping the line for retraining. For a precision task such as USB insertion, the system achieved 100 percent success, and when lighting shifted completely, the rate of human intervention was only 1.32 percent, less than one-tenth of the old method. RoHIL relights the world model, achieving 100 percent cross-lighting deployment success. EvoHIL adds adaptive capabilities to the reward model, action generator and visual perception, allowing robots to adapt to new workstations and lighting and generate new action plans without offline relabeling or line stoppage. In real-machine tests, success rates improved substantially across six typical industrial tasks, and combined training, transfer and reward-design costs fell by about 80 percent at the same deployment scale, according to QbitAI.

AnuVerse-0.5 is an industrial vertical model and world-action model. It uses an MoT architecture in which a video expert predicts frames and an action expert outputs actions, and is based on the 14B-parameter Helios. Its inference speed is in the hundreds of milliseconds and it can run on a desktop-level GPU, a cost and speed level that QbitAI said production-line edges can bear.

Nexus is the end-to-end deployment system. It unifies scheduling between different models and different robots, allowing VLA models and world-action models to be deployed quickly. QbitAI reported that Nexus offers fast deployment, autonomous scheduling and resource aggregation. It includes an action-level safety shield to reduce risks from incorrect actions on a production line. It has also connected with SAP EWM and Joule so robots can be integrated into a factory's existing production management system. Operational data from robots can flow back to the data foundation for the next training cycle, linking Helios and SimLab for data generation, ActiWorld for prediction, RoHIL and EvoHIL for skill evolution, and AnuVerse-0.5 for execution.

QbitAI reported that Anu's development path differs from the common approach of polishing a foundation model first and then finding a scenario. The company sent engineers into the real production line of Fulin Precision before building a strong base model. Anu Intelligence chairman and CEO Wen Hongjie said that investors often look first at technology route, team background, market space and capital efficiency, but his understanding changed after entering the factory. He said model capabilities will continue to spill over across the industry, and good performance on a single technical metric on a real production line cannot guarantee that a whole system will run reliably over time. Deployment experience, he said, accumulates and becomes a competitive weapon. Anu built its cognitive base and action model in a few months based on open-source models, according to the report, but deployment experience on production lines must be accumulated firsthand.

QbitAI also cited a global embodied AI research report released by Goldman Sachs at the end of August. The report said robot walking and navigation can largely be solved through simulation training, but tasks such as flexible object manipulation, tool use and fine grasping "essentially cannot be simulated."