There is more variety in training.

This year, GPU expert NVIDIA introduced GR00T N1, an open foundation model for humanoid robots. Its architecture comprises a dual system: vision-language for perception and language comprehension, and a "diffusion transformer" for generating fluid motor actions in real time. Training for GR00T N1 can be as varied as that of a top athlete. In addition to real robot trajectories, it can include videos of humans and synthetic data. According to NVIDIA, GR00T N1 should perform better than previous models in simulations and real-world tasks.

This robot can handle a little grass.

Multimodal AI enables continuous optimization, especially with regard to the movement patterns of humanoid robots designed for industrial use. A few weeks ago in Shanghai, Kepler Robotics Co., Ltd., announced the completion of significant improvements to the K2 "Bumblebee" humanoid robot's gait and released China's first demonstration video of a "hybrid, disturbance-resistant" gait. The K2 "Bumblebee" can now confidently move across various complex surfaces, including bricks, plastic, and grass. It can also maintain a stable gait when subjected to external impacts, demonstrating its ability to balance and adapt to real-world environments. Kepler also presented a demonstration of its VLA+ training platform, showing how the robot uses a multi-layered VLA+ model to interpret natural language commands and perform various tasks. These improvements in mobility and cognition position the K2 "Bumblebee" for future use in industrial applications, service scenarios, and a growing number of real-world applications.

Robots are preparing for the ultimate challenge of customer interaction.

In particular, China plans to make greater use of humanoid systems in retail, logistics, and services. Virtual Layout Assistant (VLA) models play a major role in this context because they are suitable for tasks that cannot be precisely preprogrammed. These tasks include mixed product handling, cleaning, and the supreme discipline of customer interaction. However, despite the great progress that has been made, there are still hurdles to overcome. Thus far, the most impressive performances of VLA models have been in simulations or controlled environments. Even though the challenges are significant in highly complex factories and even more so in chaotic households, 2025 could be the year that VLA models truly become practical — not only in research laboratories but also in manufacturing, warehouses, service areas, and households. Companies offering robust, reliable, and affordable solutions will prevail. As costs and installation complexity continue to decline, service and assistance robots, touted since the 1950s, will finally become a reality in private households.