At the 2026 World Robot Expo, held in Beijing from August 19-23, a dedicated forum explored how AI large models are reshaping the robotics and embodied intelligence landscape. Luchen Technology Vice President Zhao Yifei took the stage to present his vision for the future, arguing that the next generation of embodied intelligence requires a fundamentally new type of training infrastructure. He framed this as building the "operating system" for world models, a critical layer for enabling robots to learn and adapt in increasingly complex environments.

Zhao began by outlining Luchen Technology's journey since its founding in 2021, a period marked by the rapid evolution of AI infrastructure. He observed that the industry has moved from a training-centric approach to one focused on post-training and inference optimization, ultimately leading to today's environment of massive token generation. He asserted that embodied intelligence and world models are now spearheading a new phase of scaling, but this phase presents unique bottlenecks compared to large language models: the scarcity of physical world data and the challenge of training efficiency.

He elaborated on the data bottleneck, contrasting the internet's trillions of tokens of text and images with the limited availability of real-world robot interaction data. For a robot to learn a simple action like grasping or plugging, it requires real equipment, real scenarios, and real time. Even a significant dataset like Open X-Embodiment, which pools data from multiple global research institutions, contains only around one million real robot trajectories. This creates a critical mismatch: the pace of data generation in the physical world cannot keep up with the speed of model scaling, making the development of world models increasingly important.

To simplify the discussion, Zhao categorized current robot learning approaches into three distinct paths. The first is imitation learning, where robots learn from successful demonstrations, a method exemplified by models like RT-2 and OpenVLA. While effective, this approach has an inherent flaw: it primarily teaches robots the "correct answer" rather than how to recover from mistakes. This can lead to distributional shift problems, where small positional errors compound and cause task failure if not corrected. The second path, real-world reinforcement learning, allows robots to learn from trial and error, but it is expensive, slow, and fraught with equipment wear and safety risks.

This brings the discussion to the third and increasingly crucial path: learning a world model to create a virtual practice environment for robots. Instead of always failing in the real world, a robot can generate numerous "imagined rollouts" within this model to optimize its policy, then transfer these learned skills to physical hardware. As Zhao put it, this approach lets robots "practice extensively in a dream state before stepping into reality." He highlighted a significant trend where video generation models are being adapted into video-based world models for robots. This is a natural progression because a robust video model already learns to predict the next state of the world given the current one; a world model extends this to predict outcomes based on specific actions.

Zhao identified two notable examples where teams are building upon the open-source video generation model OpenSora, which was originally created by the Luchen team. The first is Microsoft Research's IGOR, which tackles the challenge of learning from massive amounts of internet videos that lack robot action labels. The team constructed a unified latent action space to align robot and human activity videos, pre-training on a dataset of roughly 2.8 million trajectories and video clips. Zhao found it particularly insightful that a team like Microsoft Research chose to build its foundation world model from the pre-trained Open-Sora, adapting it for robot control rather than starting from scratch, thus validating the value of a mature open-source base.

The second case study Zhao presented was WMPO, a collaborative project between the Hong Kong University of Science and Technology and ByteDance Seed. This work directly addresses how robots can learn to self-correct after making mistakes. The process involves training a video generation world model, having a Vision-Language-Action (VLA) policy generate numerous imagined trajectories within it, and using a reward model to judge success or failure. The policy is then updated via reinforcement learning with GRPO. The results were compelling: success rates on MimicGen's four robot tasks improved from 33.6% to 57.6%, and on a real-robot high-precision plugging task, from 53% to 70%. More importantly, the robot began exhibiting self-correction behaviors, such as re-lifting and re-aligning a plug after a missed insertion, learning these recovery skills through trial and error in the world model rather than from explicit human demonstrations. Notably, the video world model backbone for WMPO is also inherited from OpenSora.

Zhao then turned to the new demands scaling world models place on the underlying infrastructure. He highlighted three key areas. First, models are becoming increasingly heavy. Video data is multi-dimensional, requiring architectures like DiT or spatial-temporal Transformers, which impose significant pressure on GPU memory, computation, and communication. This demands robust tensor and sequence parallelism, capabilities that Colossal-AI was designed to provide. Second, maintaining optimization stability during large-scale training is challenging. As global batch sizes expand, hyperparameters like learning rate and warm-up schedules must be carefully managed. Zhao cited the LAMB optimizer, co-created by his company's founder Yang You in 2019, which dramatically accelerated BERT training by enabling stable, large-batch optimization. He noted that this capability is already being applied in robotics, referencing the SpatialActor model, an AAAI 2026 Oral paper, which uses LAMB across over 50 robot manipulation tasks. Third, Zhao pointed to the unique "train-inference alternation" pattern of world models. In systems like WMPO, the world model generates trajectories (inference), the reward model evaluates them (inference), and the VLA policy is updated (training). As the policy changes, new behavior data can further fine-tune the world model. This creates an iterative loop where training and inference are no longer separate phases but a continuous, integrated process, demanding infrastructure that can flexibly schedule both.

Looking ahead, Zhao proposed a four-layer architecture for the future of world model infrastructure. The foundation is compute resources and cloud infrastructure. The next layer is Colossal-AI for solving large-scale parallel computing, memory, and communication challenges. Above that sits LAMB and similar optimization algorithms to ensure training efficiency and stability at scale. At the top is OpenSora, providing a validated technical starting point for video generation models. He believes this combination forms the true "operating system" for world models.

Zhao concluded by summarizing the value his company aims to deliver to embodied intelligence teams in four key points: a higher starting point through the use of mature open-source models; faster and more stable development through joint algorithm and infrastructure optimization; lower costs by minimizing GPU hours and heavy capital investment; and a complete closed loop by connecting world model pre-training with policy learning, evaluation, and deployment. He drew a parallel to the PC and internet industries, where standardized infrastructure enabled rapid innovation at the application layer. Just as no startup would build its own Linux or database, he argued that robotics companies should focus on their core competencies—robot hardware, application scenarios, proprietary data, and algorithm innovation—while leaving complex, common training infrastructure to specialized platforms like Luchen Technology. He ended with the vision of enabling robots to learn "in a dream" more cheaply and efficiently, ultimately preparing them to step into the real world.