Scalabot HERON-CRA Adds Context Memory and RL Engine for Embodied Control

Songyan Dynamics has launched the HERON-CRA model, the second in its HERON stack following the HERON-World Model. HERON-CRA aims to assist robots in remembering task history, correcting mistakes, and transferring skills across different robot bodies. The model comprises three main components, with the Context Expert encoding multimodal history as structured spatiotemporal tokens, allowing policies to condition on more than the current frame. The model shows a significant performance improvement, with success in a sock-folding task rising from 38.5% to 97.8% when the RL Engine is activated.