Yuanli Lingji has released DW0.5, a method that uses world models to train VLA (Vision-Language-Action) models. By moving reinforcement learning (RL) into a virtual world, the approach reduces the need for real-world data by 60% during post-training. This could accelerate robot learning and reduce dependence on physical data.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Yuanli Lingji releases the DW0.5 model, which uses a world model to provide a virtual training environment for VLA, shifting reinforcement learning from the physical world to the virtual world, reducing post-training real robot data requirements by 60%. This breakthrough could reshape the training paradigm for embodied intelligence.
The core innovation of Yuanli Lingji's DW0.5 model is using a world model as a 'coach' to provide a virtual training environment for Vision-Language-Action (VLA) models. Traditional VLA training relies on large amounts of real robot data, which is costly and difficult to scale. DW0.5 builds an interactive virtual world where VLA models can perform reinforcement learning (RL), thereby reducing dependence on physical robot data.
According to Yuanli Lingji's official data, after adopting DW0.5, post-training real robot data requirements drop by 60%. This means that data that originally required 100 hours of real robot collection now only needs 40 hours, significantly lowering the training threshold. This figure comes from internal company tests and has not been independently verified by third parties.
DW0.5's technical route is to transfer RL training from the physical world to the virtual world. Specifically, the world model learns the dynamics of the environment, including object interactions, physical laws, etc., and then provides simulated perception inputs and action feedback for the VLA model. The VLA model learns strategies through trial and error in the virtual environment, and then transfers the learned strategies to the real robot.
The key to this method is the fidelity of the world model. If the virtual environment differs too much from the real environment, the transfer effect will be greatly reduced. Yuanli Lingji states that DW0.5, trained on multimodal data, can capture fine-grained physical interactions, but specific evaluation metrics (such as Sim-to-Real success rate) have not been disclosed.
The release of DW0.5 has potentially transformative significance for the field of embodied intelligence. Real robot data collection has always been a bottleneck in robot learning: high cost, long cycle, and limited scenarios. If world models can effectively replace part of the real robot data, it will accelerate the transition of robots from the lab to practical applications.
However, the technology still faces challenges. First, the generalization ability of world models in complex dynamic scenes (such as human interaction, unstructured environments) has not been verified. Second, the 60% data reduction was measured in specific tasks (such as tabletop manipulation), and whether it can be extended to a wider range of tasks remains to be seen. In addition, Yuanli Lingji has not disclosed details such as the model size and training data volume of DW0.5, making it difficult for outsiders to assess its reproducibility.
Yuanli Lingji is not the only company exploring the use of world models for robot training. Google DeepMind's RT-2 and Stanford's VoxPoser also attempt to use vision-language models or world models to reduce real robot data requirements. DW0.5's uniqueness lies in explicitly placing RL training entirely in the virtual world, rather than just using world models for data augmentation.
In the future, if DW0.5 can verify its effectiveness on more tasks and open-source or provide an API, it may drive a shift in industry standards. But in the short term, real robot data remains indispensable, and world models are more of a supplementary tool. Yuanli Lingji plans to improve the real-time performance and fidelity of the world model in subsequent versions and explore cooperation with hardware manufacturers.
The information in this article mainly comes from Yuanli Lingji's official release, which is corporate promotional material. Key data (60% reduction in real robot data requirements) has not been verified by third parties, and detailed experimental settings have not been provided. Readers should view it with caution and wait for subsequent independent evaluations.
Yuanli Lingji's DW0.5 moves RL training into the virtual world via a world model, reducing post-training real robot data needs by 60%, but the technology's maturity and generalization ability remain to be verified.
Primary report
Primary source