PANews reported on January 2nd that Andrew Kang, a partner at Mechanism Capital, stated in an article published on the X platform that by 2025, the robotics field will have resolved long-standing challenges in model architecture and training, and made significant progress in data acquisition technology, data quality understanding, and data formulation. This will give AI companies the confidence to finally begin investing in large-scale data collection. Companies like Figure, Dyna, and PI have achieved success rates of over 99% in various real-world application scenarios by leveraging innovative reinforcement learning (RL) techniques.

Furthermore, advancements in memory technology have broken down the "memory wall." NVIDIA's ReMember utilizes memory-based navigation, while Titans and MIRAS achieve test-time memory. A superior Virtual Localization Model (VLM) means that Virtual Localization Arrays (VLAs) possess better spatial understanding capabilities, as well as data annotation and processing workflows that can significantly improve throughput. By 2025, the market will begin to experience the zero-shot capability mapping, visual strength sensitivity, and general physics reasoning brought about by the scale of data. By 2026, the scale of entity AI data will expand 100 times.