This work presents a training procedure and a pretrained model that can estimate the normal force distribution and the contact area acting upon the GelSight Mini visuotactile sensor. The representation maps from the raw output images of the sensor to the normal force distribution acting across the entire sensor surface. This is appealing as it allows to directly do classical force control based on the output of visuotactile sensors. Moreover, it yields a physically grounded, interpretable representation of the tactile signals.

This work investigates the effectiveness of tactile sensing for the practical everyday problem of stable object placement on flat surfaces starting from unknown initial poses. We devise a neural architecture that estimates a rotation matrix, resulting in a corrective gripper movement that aligns the object with the placing surface for the subsequent object manipulation. We compare models with different sensing modalities, such as force-torque, an external motion capture system, and two classical baseline models in real-world object placing tasks with different objects. The experimental evaluation of our placing policies with a set of unseen everyday objects reveals significant generalization of our proposed pipeline, suggesting that tactile sensing plays a vital role in the intrinsic understanding of robotic dexterous object manipulation.

We propose learning task-space, data-driven cost functions as diffusion models. Diffusion models represent expressive multimodal distributions and exhibit proper gradients over the entire space. We exploit these properties for motion optimization by integrating the learned cost functions with other costs in a single objective function, and optimize all of them jointly by gradient descent.

We propose a novel hybrid method for Robot Assembly discovery that is based on a combination of Mixed Integer Programming and a graph-based reinforcement learning agent.

We propose a novel method for learning to assemble arbitrary structures from scratch. The transformer-like graph-based neural network jointly decides which blocks to use and how to assemble the structure with the robot-in-the-loop.

We present a benchmark of structured policies for real-world dexterous manipulation. Our proposed approaches combine elements of classical robotics and modern policy optimization. This inclusion of inductive biases facilitates sample efficiency, interpretability, reliability and high performance.