Songyan Dynamics has released HERON-CRA (Context Reinforcement Action Model), the second model in its HERON series, following the presentation of HERON-World Model last week. The company positions HERON-CRA as a foundational model that integrates context, reinforcement learning, and actions, designed to assist robots in remembering task history, correcting errors, and transferring skills between different robots.
The release is defined by three core components. A Context Expert encodes multimodal history into structured 4D spatio-temporal tokens, allowing the policy to consider not only the current frame but also memory windows exceeding one minute. The Reinforcement Learning Engine functions as a lightweight residual actor-critic, reusing frozen keys and values from the Context Expert to learn corrective actions when the simulation itself begins to deviate.
Cross-embodiment pretraining combines teleoperation, simulation, UMI-style data, and first-person video collected from various manipulators, wheeled platforms, grippers, and dexterous hands. This enables the model to be applied to new morphologies after training with less need for fine-tuning.
Demonstrations highlight the claimed performance gains. In a soft object sock folding task, Scalabot reports an increase in success rate from 38.5% using only simulation to 97.8% after activating the RL Engine. Individual videos show a long-horizon coffee preparation sequence—from grinding to pouring milk—completed in under 30 seconds. Furthermore, the transfer of this skill set between the ARX manipulator and the Noetix M1 wheeled humanoid was demonstrated. The model is also capable of recovering if a cup moves during grasping and continues some movements during brief occlusion.
Technically, the Context Expert and Action Expert are trained as a mixture of transformers. The reinforcement learning path adds a value model with time difference constraints and can play out imagined trajectories through the HERON-World Model without additional hardware risks. While independent teams may wish to reproduce the sock folding and coffee metrics on their own equipment, this release provides a clear algorithm—memory, residual RL, and cross-embodiment pretraining—for performing long-horizon domestic manipulations beyond single-step demonstrations.