A common idea that was in the air in the final conference of the Sharework project “Unlocking industrial human-robot collaboration” held in Barcelona on October 19, 2022, was that data-driven methods are starting to emerge as novel tools that will change the robotics paradigm as we know it. Nevertheless, Jan Peters, Professor at TU Darmstadt and leader of the Intelligent Autonomous Systems lab, claimed that this integration should not be done naively, but rather, taking into account the inductive biases of the problem we aim to solve reducing the search space of the learning methods.
These inductive biases have been clear for image or text models. Convolutional networks can extract the spatial features of an image, while transformers or Recurrent Neural Networks encode the temporal information of the text. The choice of these networks highly improves the performance of the learned models in contrast to using naively Fully Connected Neural Networks. But, what are the inductive biases for Robotics? How should we model robot skills?
Geometric Inductive Biases in Robotics. Learning 6D robot pose generative models.
In our recently submitted paper to ICRA 2023, we study the problem of learning generative models for robot grasp poses as diffusion models. The learned model should adapt to multiple objects to grasp and generate a set of valid grasps.
To enhance the performance of the model, we assumed multiple inductive biases. We transformed the end-effector pose to a set of 3D point by a geometry encoder, considered rotation invariant Pointcloud encoders and learned the SDF of the graspable object as shown in the architecture below.
As shown in the following image, the model was able to generate grasp poses for a whole myriad of objects.
Julen Urain is a Ph.D student in the Intelligent Autonomous Systems department in Technical University Darmstadt (TUDA) advised by Jan Peters. He holds a master in Robotics and Control, with a Master thesis in Robot-Human Interaction. His research interests are in the interplay of Geometry, Optimisation and Deep Learning for reactive motion generation in robots.
The Technische Universität Darmstadt is one of Germany’s leading technical universities with 23,000 students and 270 professors. The Intelligent Autonomous Systems Institute, one of the strongest robotic learning groups in Europe, contributes to the project leading the development of learning algorithms and the study and formalisation of human gestures.