Discovering the symbols and rules that can be used in long-horizon planning from a robot's unsupervised exploration of its environment continuous sensorimotor experience is challenging task. The previous studies proposed learning single or paired object interactions with these symbols. In this work, we propose system learns discovered relational encode an arbitrary number objects relations between them, converts those to Planning Domain Description Language (PDDL), generates plans involve...

Learning to interact with the environment not only empowers agent manipulation capability but also generates information facilitate building of action understanding and imitation capabilities. This seems be a strategy adopted by biological systems, in particular primates, as evidenced existence mirror neurons that seem involved multi-modal understanding. How benefit from interaction experience robots enable actions goals other agents is still challenging question. In this study, we propose...

Symbolic planning and reasoning are powerful tools for robots tackling complex tasks. However, the need to manually design symbols restrict their applicability, especially that expected act in open-ended environments. Therefore symbol formation rule extraction should be considered part of robot learning, which, when done properly, will offer scalability, flexibility, robustness. Towards this goal, we propose a novel general method finds action-grounded, discrete object effect categories...

We propose a novel general method that finds action-grounded, discrete object and effect categories builds probabilistic rules over them for non-trivial action planning. Our robot interacts with objects using an initial repertoire is assumed to be acquired earlier observes the effects it can create in environment. To form action-grounded object, effect, relational categories, we employ binary bottleneck layer predictive, deep encoder-decoder network takes image of scene applied as input,...

Exploration and self-observation are key mechanisms of infant sensorimotor development. These processes further guided by parental scaffolding to accelerate skill knowledge acquisition. In developmental robotics, this approach has been adopted often having a human acting as the source scaffolding. study, we investigate whether Large Language Models (LLMs) can act agent for robotic system that aims learn predict effects its actions. To end, an object manipulation setup is considered where one...

Abstraction is an important aspect of intelligence which enables agents to construct robust representations for effective decision making. In the last decade, deep networks are proven be due their ability form increasingly complex abstractions. However, these abstractions distributed over many neurons, making re-use a learned skill costly. Previous work either enforced formation creating designer bias, or used large number neural units without investigating how obtain high-level features...

In this paper, we propose a concept learning architecture that enables robot to build symbols through self-exploration by interacting with varying number of objects. Our aim is allow learn concepts without constraints, such as fixed interacted objects or pre-defined symbolic structures. As such, the sought should be able for single can grasped, object stacks cannot grasped together, other composite dynamic Towards end, novel architecture, self-attentive predictive encoder-decoder network...

In this paper, we propose a graph-based image-to-image translation framework for generating images. We use rich data collected from the popular creativity platform Artbreeder <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> , where users interpolate multiple GAN-generated images to create artworks. This unique approach of creating new leads tree-like structure one can track historical about creation particular image. Inspired by...

Discovering the symbols and rules that can be used in long-horizon planning from a robot's unsupervised exploration of its environment continuous sensorimotor experience is challenging task. The previous studies proposed learning single or paired object interactions with these symbols. In this work, we propose system learns discovered relational encode an arbitrary number objects relations between them, converts those to Planning Domain Description Language (PDDL), generates plans involve...

In this paper, we propose and realize a new deep learning architecture for discovering symbolic representations objects their relations based on the self-supervised continuous interaction of manipulator robot with multiple in tabletop environment. The key feature model is that it can take changing number as input map object-object into domain explicitly. model, employ self-attention layer computes discrete attention weights from object features, which are treated relational symbols between...

Generative adversarial networks (GANs) are deep neural that allow us to sample from an arbitrary probability distribution without explicitly estimating the distribution. There is a generator takes latent vector as input and transforms it into valid also discriminator trained discriminate such fake samples true of distribution; at same time, generate fakes cannot tell apart samples. Instead learning global generator, recent approach involves training multiple generators each responsible one...

Recent works show that learning contextualized embeddings for words is beneficial downstream tasks. BERT one successful example of this approach. It learns by solving two tasks, which are masked language model (masked LM) and the next sentence prediction (NSP). The pre-training can also be framed as a multitask problem. In work, we adopt hierarchical approaches pre-training. Pre-training tasks solved at different layers instead last layer, information from NSP task transferred to LM task....

In this paper, we propose a graph-based image-to-image translation framework for generating images. We use rich data collected from the popular creativity platform Artbreeder (http://artbreeder.com), where users interpolate multiple GAN-generated images to create artworks. This unique approach of creating new leads tree-like structure one can track historical about creation particular image. Inspired by structure, novel graph-to-image model called Graph2Pix, which takes graph and...

This paper proposes an architecture that can learn symbolic representations from the continuous sensorimotor experience of a robot interacting with varying number objects. Unlike previous works, this work aims to remove constraints on learned symbols such as fixed interacted objects or pre-defined structures. The proposed for single and relations between them in unified manner. is encoder-decoder network binary activation layer followed by self-attention layers. Experiments are conducted...

Exploratoration and self-observation are key mechanisms of infant sensorimotor development. These processes further guided by parental scaffolding accelerating skill knowledge acquisition. In developmental robotics, this approach has been adopted often having a human acting as the source scaffolding. study, we investigate whether Large Language Models (LLMs) can act agent for robotic system that aims to learn predict effects its actions. To end, an object manipulation setup is considered...

In this paper, we propose and realize a new deep learning architecture for discovering symbolic representations objects their relations based on the self-supervised continuous interaction of manipulator robot with multiple tabletop environment. The key feature model is that it can handle changing number naturally map object-object into domain explicitly. model, employ self-attention layer computes discrete attention weights from object features, which are treated as relational symbols...

Learning to interact with the environment not only empowers agent manipulation capability but also generates information facilitate building of action understanding and imitation capabilities. This seems be a strategy adopted by biological systems, in particular primates, as evidenced existence mirror neurons that seem involved multi-modal understanding. How benefit from interaction experience robots enable actions goals other agents is still challenging question. In this study, we propose...