Panasonic HD develops "Diffusion Contact Model" AI technology for robot control that applies generative AI to perform contact-rich actions

To enable a robot to perform safe and accurate movements, the parameters that control the robot's movements and the amount of force must be finely tuned to suit the situation the robot is in. Tuning methods are mainly divided into simulation environments (model-based) and trial and error using the actual machine (machine-based), with the model-based approach featuring the advantage of being able to carry out a large amount of trial-and-error simulations more efficiently than a machine-based approach. However, tasks that involve contact are difficult to simulate due to the complexity of the movements and forces (contact dynamics) that occur when a robot touches a person or object, and it is necessary to carry out trial and error many times when training a robot for these tasks, and the training also must involve human input.
 
Therefore, Panasonic HD focused on the features of diffusion models now being used for image and sound generation, as they can express complex, nonlinear models.

Applying the similarity between the noise removal process of the diffusion model and the optimization process of contact simulation, Panasonic HD developed its Diffusion Contact Model that can simulate complex contact dynamics without using an actual machine. The Diffusion Contact Model simulates in stages the force exerted when a robot touches an object , and can predict the force exerted when the robot touches an object with high accuracy, enabling efficient tuning of control parameters on a model basis.

In the conventional method shown in the upper portion of Figure 1, the control parameters are first estimated using a Bayesian optimization algorithm, then evaluated on an actual machine and the parameters are adjusted again, repeating a trial-and-error loop until the desired performance is obtained. On the other hand, in the Diffusion Contact Model shown in the lower portion of the figure, the control parameters are first estimated using a Bayesian optimization algorithm as in the conventional method, but the Diffusion Contact Model is used instead of an actual machine for evaluation.

First, Panasonic HD conducted experiments in a simulation environment, and found that the Diffusion Contact Model can predict contact forces with high accuracy compared to a conventional deep neural network (DNN). If complex contact dynamics can be simulated without an actual machine, the number of situations requiring an actual machine can be reduced compared to conventional methods. In a demonstration using an actual machine for a wiping task, it was shown that the learning time for the wiping task, which took 80 minutes in total by mainly teaching the robot by hand, can be reduced to about 25 minutes using the Diffusion Contact Model.