Star Dust has released the online reinforcement learning framework SmoothRL, enabling asynchronous inference with large models.

On September 4, 2026, the Astribot base model team officially released the SmoothRL online reinforcement learning framework. This framework aims to address the issue of continuously improving robot capabilities from actions after asynchronous inference becomes the norm in large models, making robots more precise, stable, and reliable. SmoothRL supports the continuous progress of reinforcement learning during model inference, without waiting for the inference to complete. For the first time, asynchronous online reinforcement learning was applied to real high-dynamic throwing tasks, demonstrating that this technology can be used not only for high-precision correction but also for dynamic operations such as continuous acceleration and precise release where pause is impossible. Astribot is leveraging its “AI model - embodied OS - rope-driven body” full-stack system to promote the application and large-scale deployment of Physical AI. Next, they plan to explore broader strategy updates, more extensive task distributions, and the combination of asynchronous execution with end-to-end optimization of generative strategies.