X Square Robot has released HOST, an open-source inference-time learning framework that allows a humanoid robot to watch a 29-second human demonstration video and then attempt to reproduce the shown skill.
The system operates by processing the demonstration at inference time rather than requiring offline fine-tuning on large datasets, representing a shift from traditional embodied AI approaches.
In testing, the robot achieved a 62 percent success rate when replicating tasks after viewing the short video demonstrations.
The framework is designed to enable on-the-fly imitation, allowing robots to acquire new capabilities without extensive retraining cycles.
By open-sourcing HOST, X Square Robot makes the inference-time learning approach available for further research and development by the broader robotics community.
The technology addresses a key challenge in robotics: reducing the time and data required for robots to learn new manipulation skills from human examples.
Researchers note that the 62 percent success rate indicates both progress and room for improvement in single-shot visual imitation for complex humanoid control.
The open-source release includes the framework code and documentation to facilitate replication and extension of the inference-time learning method.
A Robot That Learns from Short Videos in 29 Seconds — X Square Robot's HOST Changes the Embodied-AI Recipe
This is an independent summary. The complete reporting, supporting context and any primary documents remain with Pandaily.