On January 6, 2026, Zhiyuan's Embodied Research Center made a significant announcement by introducing the SOP (Scalable Online Post-training) system. This innovative system marks a groundbreaking achievement as it, for the very first time, seamlessly integrates online learning, distributed architecture, and multi-task generalist capabilities into post-training for physical-world VLA (Vision-Language-Action).

The SOP system utilizes an Actor-Learner asynchronous architecture. This architecture empowers multiple robots to gather data simultaneously and then upload it to the cloud. Moreover, it has the capability to dynamically adjust the ratio of online to offline data. As a result, it can achieve model synchronization and updates within minutes.

Experimental results are highly promising. When the HG-Dagger method is combined with the SOP system, there is a remarkable 33% improvement in performance within supermarket scenarios. In terms of clothing-folding tasks, the throughput increases by an impressive 114%. Additionally, the system achieves multi-task success rates that surpass 94%.

When it comes to training efficiency, a four-robot cluster trained using the SOP system demonstrates a training speed that is 2.4 times faster compared to a single machine. This clearly validates the effectiveness of scalable learning.

In essence, the SOP system enables robots to continuously evolve during real-world deployments, effectively reshaping the entire lifecycle of robots.