The Massachusetts Institute of Technology (MIT) showcased a brand new robot training model this week that abandons the previous focus on training methods centered around specific datasets, opting instead for an approach that utilizes vast amounts of information similar to that used in training large language models (LLMs). Researchers noted that imitation learning—where agents learn by mimicking individuals performing tasks—can fail when facing minor challenges. These challenges may include varying lighting conditions, different environmental setups, or new obstacles. In such cases, the robots lack sufficient data to adapt to these changes.

Norwegian startup 1X Technologies recently announced significant advancements in developing AI-based world models for robots. Simply put, these models act as virtual testing grounds for robots, allowing them to test and improve in various scenarios without the need for real-world testing. 1X believes this is key to solving the 'robotics challenge' — how to reliably evaluate robots trained for multiple tasks in ever-changing environments.

Research shows that the MimicGen system can generate large-scale training data for robots. By leveraging digital twin technology, it replicates human operation data in a simulated environment. Using fewer than 200 human demonstrations, it autonomously generates 50,000 training data points. Synthetic data and simulation are crucial for AI development, providing an infinite source of training data. MimicGen is suitable for long-sequence tasks and high-precision tasks, such as grasping and assembling.