LightX Innovation has officially released its full-body intelligence foundation model, Light-O1, designed to learn human actions from internet videos and integrate linguistic and visual cues to transfer this knowledge to robots, thereby enhancing their ability to coordinate full-body movements and complete complex tasks. This marks a significant step forward in bridging the gap between human behavior and robotic execution capabilities.

A key breakthrough of Light-O1 is its initial validation of the cross-platform transfer scaling law for full-body human motion pre-training. The model demonstrates that as the volume of human action data used for pre-training increases, the accuracy of action prediction improves when the model is adapted to different robotic platforms and viewpoints. This empirical finding provides a robust foundation for leveraging massive-scale human data to advance the general-purpose capabilities of robots.

Following the company's previous successes in Scalable Alignment and Scalable Deployment, Light-O1 propels the Scalable Pre-Training phase of LightX Innovation's Physical AI three-stage paradigm into practical application. This development represents a critical evolution from limited, labor-intensive data collection to harnessing the vast and diverse repository of human behavior available online.

Currently, robotic action models primarily depend on specialized data collection methods, such as teleoperation or Universal Manipulation Interface (UMI). These approaches are costly, time-consuming, and require custom processes, personnel training, and dedicated equipment for each new task, leading to significant logistical challenges as tasks multiply. These constraints have historically made it difficult to scale both the volume and diversity of data concurrently, posing a major bottleneck for the scaling of robot foundation models. Light-O1 overcomes this by utilizing the immense amount of human behavior already captured in internet videos, which offers an unparalleled scale and variety of actions, including moving, manipulating objects, using tools, and interacting with others, thereby supplementing and expanding beyond traditional data collection methods.

To achieve this, Light-O1 establishes a comprehensive technical pipeline that covers data construction, autoregressive pre-training, and robotic adaptation. The core strategy involves learning reusable action knowledge from rich human behaviors and transferring it to robots that differ in body structure, observation perspective, and control methods. The process begins by transforming video-based human actions into unified multimodal data, then uses large-scale pre-training to develop action priors—understanding what actions to take in a given context—and finally, adapts these priors using target robot data to convert human behavioral experience into actionable robotic capabilities.

Through this pipeline, LightX Innovation conducted experiments using varying sizes of video-based human action data for pre-training. The models were then adapted to first-person human data, public humanoid robot data, and LightX's proprietary robot data for full-body mobile manipulation. The results conclusively demonstrated the cross-robot transfer scaling law: with an expansion of pre-training data to 100,000 action hours, the action prediction loss followed a power-law decline, and full-body pose errors in offline open-loop evaluations were correspondingly reduced. This confirms that the benefits of human motion pre-training can successfully migrate to robot action prediction, regardless of the differences in physical platforms, and validates the rationale for scaling up human action data to enhance general-purpose robotic abilities.

Light-O1 connects instruction understanding, environmental perception, and full-body action to demonstrate universal full-body intelligence in real-world scenarios. In physical environments, the self-developed small humanoid robot LightBot, equipped with Light-O1, can autonomously perform tasks such as handing over towels, opening shoe cabinets to organize slippers, and picking up trash. It successfully integrates action planning, mobile navigation, and whole-body manipulation into a seamless process. The robot shows remarkable resilience, adjusting its actions based on environmental feedback and execution results, and autonomously retrying when faced with failures, external interference, or changes in object types and positions.

Furthermore, Light-O1 is capable of understanding natural language instructions, inferring the user's needs and implicit intentions, and reasoning out the necessary body posture, action sequences, and constraints to generate corresponding full-body actions. Coinciding with this release, LightX Innovation has open-sourced Light-O1-Preview, a general-purpose action generation model on its official website. This allows users to describe a desired outcome and have the model deduce and display the corresponding body movements and reasoning process, offering a direct experience of the entire journey from understanding a request to generating an action.

With this launch, LightX Innovation has now secured core technological achievements for all three stages of its Physical AI paradigm: Scalable Pre-Training, Scalable Alignment, and Scalable Deployment. This forms a complete technical chain from foundation model pre-training to capability alignment and real-world deployment. Crucially, this framework establishes a sustainable growth path for Physical AI capabilities: base abilities are continuously accumulated through scalable pre-training, their integration into the physical world is enhanced through scalable alignment, and real deployment introduces the models to an ever-expanding array of tasks, environments, and robotic forms. This interconnected approach ensures that Physical AI development is not reliant on isolated breakthroughs but is instead driven by a systematic, iterative path from data and training to application, fostering continuous improvement and broader impact.