Hedra Unveils Model Factory for Training and Enhancing AI Models in Robotics

The American technology company Hedra introduced the Model Factory platform on November 10, 2026, designed for creating specialized AI models based on existing solutions. It integrates computational infrastructure, additional training, performance optimization, and tools for continuous model improvement after deployment. The service is intended for developers in robotics, industrial systems, biological research, and defense technologies.

The main feature of Model Factory is the application of the Recursive Self-Improvement (RSI) approach. It involves a sequential cycle of experiments, result evaluation, and model adjustments using autonomous AI agents. The process remains under the control of researchers who set training goals and verify the quality of the solutions obtained.

How Hedra Model Factory Works

To adapt a model, the customer provides a basic AI system and data related to a specific task. This could include video recordings, sensor information, computer simulation results, or industrial equipment parameters. Hedra conducts computations in an isolated GPU environment, after which company specialists, together with AI agents, organize additional training.

Multimodal models capable of analyzing images, videos, and other types of information are used in the process. They help evaluate training data, provide feedback, and conduct experiments with various settings. This approach allows the adaptation of an existing model to specific conditions without creating a new system from scratch.

After training is completed, the customer can deploy the model on their own servers or in Hedra's partner data center. For devices with limited computational resources, model optimization through distillation is provided—a method of reducing computational complexity while maintaining necessary capabilities.

The developer also suggests using data obtained after model deployment for further improvement. Researchers analyze the results and organize new training cycles, with updated versions released after passing checks established by the customer.

What Results the Technology Showed

As an example, Hedra presented the optimization of the Krea 3 model, designed for generating and editing images. According to the company, the image processing time at 2K resolution was reduced from approximately 22 to 3.6 seconds, roughly six times faster. The work took less than a week and was conducted on the same hardware configuration of four AMD Instinct MI355X accelerators.

The acceleration resulted from several changes, including model distillation, computation optimization, and reducing the number of generation steps from 30 to 14. Therefore, this figure cannot be considered a universal performance gain of Model Factory: it pertains to a specific model, test load, and chosen settings.

In another experiment, Hedra improved the ability of the H3 video model to reproduce physical processes. After additional training, its result in the Physics-IQ test increased from 30.3 to 34.9 points. This test, developed with the participation of Google DeepMind, evaluates how realistically generative video models depict object movement, collisions, and other physical phenomena.

The company also reported that the distilled version of the video model was able to create five-second clips in less than three seconds on AMD MI355X accelerators. However, high generation speed alone does not guarantee the physical accuracy of all created scenes.

Experiment with Robotic Grasp

A separate study by Hedra focused on adapting a robotic system after replacing the gripper—the device the robot uses to hold and move objects. Such changes can disrupt the operation of a previously trained model because the geometry and parameters of the new equipment differ from the original ones.

In the simulation, developers replaced the robot's gripper, after which the success rate of the assigned operation dropped almost to zero. The system then matched the new gripper parameters with the image from the camera mounted on the robotic arm. According to Hedra, this allowed the success rate to be restored to nearly 100% without retraining the model.

The result demonstrates the possibility of adapting AI to equipment changes by adjusting its parameter representation. However, the test was conducted in a virtual environment, so it cannot be considered proof of the method's reliability on real industrial robots.

Where Model Factory Will Be Applied

Hedra plans to use the platform for developing specialized AI systems capable of working with physical objects and processes. In robotics, such solutions can be applied for manipulator control, item sorting, and equipment adaptation to new tasks. In industry and scientific research, they can help analyze sensor data and optimize the operation of complex systems.

Hedra co-founder and CEO Michael Lingelbach, who previously worked on reinforcement learning research at Stanford University and Nvidia, believes that the next stage of AI development will be related to its application in the physical world. To this end, the company proposes combining specialized research support with automated experiments and computational infrastructure.

Hedra Model Factory is already available for corporate clients. However, the recursive self-improvement mechanisms on the company's website are designated as a research direction with preliminary access. The cost of corporate services, results of large-scale industrial implementations, and confirmed computational resource savings have not yet been disclosed.