Hedra Inc. has launched the Model Factory, a platform designed for post-training and inference of physical AI models. The service utilizes recursive self-improvement (RSI) to enable robots and machines to adapt to specific real-world tasks, such as parcel packing.
The platform functions as an autonomous training pipeline for sectors including robotics, manufacturing, biology, and defense. Customers provide base models and proprietary data, which are processed in a GPU sandbox environment. Hedra researchers use vision-language models to provide reinforcement learning feedback for continuous optimization.
The system allows for model deployment on-premises or via cloud partners. Models can be distilled for edge systems to reduce latency. Once operational, the platform creates a feedback loop where new data continuously improves the model performance.
Hedra validated the platform using the MiniMax H3 video model, improving its Physics-IQ benchmark score from 30.3 to 34.9. Additionally, the Krea 3 2K model was optimized from 30 sampling steps to 14, reducing latency from 22 seconds to 3.6 seconds in one week.
Founder Michael Lingelbach stated that the platform aims to facilitate the creation of bespoke models by reducing compute-intensive training requirements. In one demonstration, the system restored a robot's task success rate to 100% after a hardware change without retraining.
