In a significant breakthrough for the field of robotics, OpenAI-backed company 1X has developed a world model, marking the first evidence of robot scaling law. This innovative world model, based on video generation (Sora) and end-to-end autonomous driving world model (E2EAD), serves as a virtual simulator for 1X’s own robots.
Background on 1X and its World Model
1X, a Norwegian humanoid robot company founded in 2014, is dedicated to developing bipedal robots NEO and commercial wheeled humanoid robots EVE. In early this month, 1X officially launched the NEO Beta, a bipedal humanoid robot designed for home use, showcasing the continuous progress of highly realistic and super-humanoid robots powered by large models.
The world model developed by 1X is a major milestone in the company’s journey. According to 1X AI Vice President Eric Jang, over the past few months, the company has been developing a learning simulator for general-purpose robots, which has now finally been realized. 1X Engineer Daniel Ho explained that this is a preliminary progress of the robot world model, which can understand the physical world encountered and generate high-fidelity videos, allowing robots to perform planning, evaluation, and simulation operations in their own neural spaces. Many of the generated video clips are difficult to distinguish from real videos.
World Model’s Capabilities and Significance
The world model can predict the interaction of non-trivial objects (such as rigid bodies), the impact of falling objects, partial observability, deformable objects (curtains, clothes), and hinge objects (doors, drawers, curtains, chairs). This capability has led some people to compare the 1X humanoid robots to the Westworld series.
The world model is particularly valuable in simulating object interactions. In the box grasping experiment, the research team provided the model with the same initial frame and three different actions to grasp the box. In each case, the box to be grasped would be lifted and moved by the gripper, while the other boxes remained undisturbed. Even without providing actions, the world model would generate videos that seemed reasonable, such as learning to avoid people and obstacles when driving.
Robot Scaling Law and Its Importance
The robot scaling law is a key factor in the development of robots. It refers to how the capabilities of robots can expand as the size of the data, computation, and model increases. The world model developed by 1X addresses a very practical but often overlooked challenge in building general-purpose robots: evaluation.
Before the world model, 1X trained a model that saw a performance drop in 50 days. This situation makes it extremely difficult to conduct robot research in the real world. The world model enables the evaluation of general-purpose robots by directly learning from raw sensor data and using the simulator to evaluate new strategies.
Self-Cognition and Future Implications
While the world model has shown impressive capabilities, it still has limitations. For example, the model failed to recognize itself in a mirror, indicating a lack of self-awareness or self-understanding. However, this is just the beginning, and as the technology continues to evolve, we can expect to see even more advanced capabilities in robots.
The development of the world model by 1X is a significant step forward in the field of robotics. It demonstrates the potential of scaling law in building more capable and efficient robots. As the technology continues to advance, we can expect to see even more impressive achievements in the field of robotics.
Views: 0