Noetix gives its robots a memory and a way to learn from error

Noetix Robotics, a Beijing humanoid maker founded in 2023, released a second model under its embodied-intelligence brand Scalabot on 15 September 2026. HERON-CRA — Context Reinforcement Action Model — follows the HERON-World Model announced a week earlier, on 8 September.

The world model gave a robot a way to predict what a scene would look like after an action. HERON-CRA is the half that acts on that prediction, and it is built around three pieces. A Context Expert encodes what has already happened in a task into structured spatio-temporal context, so the robot is not deciding each step as though it were the first. An RL Engine adds a value model that scores progress and a lightweight residual reinforcement layer that corrects actions, trained both on real executions and on trajectories the world model simulates. Cross-embodiment pretraining uses data from different robot bodies so a skill learned on one transfers to another — humanoid, wheeled base, or a fixed manipulator.

Noetix has published no benchmark results alongside the release, and the model is described through its architecture rather than through measured task success. What the company is claiming for now is the shape of the system, not its score.