In the past few years, a considerable amount of research has been dedicated to exploitation previous learning experiences and design Few-shot Meta Learning approaches, in problem domains ranging from Computer Vision Reinforcement based control. A notable exception, where best our knowledge, little no effort made this direction is Quality-Diversity (QD) optimization. QD methods have shown be effective tools dealing with deceptive minima sparse rewards Learning. However, they remain costly due...

As open-ended learning based on divergent search algorithms such as Novelty Search (NS) draws more and attention from the research community, it is natural to expect that its application increasingly complex real-world problems will require exploration operate in higher dimensional Behavior Spaces (BSs) which not necessarily be Euclidean. traditionally relies k-nearest neighbours an archive of previously visited behavior descriptors are assumed live a Euclidean space. This problematic...

We focus on the real-time fusion of monocular visual SLAM with GPS data in order to obtain city-scale, georeferenced pose estimations and reconstructions. Recently, GPS/VSLAM through constrained local key-frame based Bundle Adjustment (BA) using Barrier Term Optimization (BTO) has proven be (to best our knowledge) most robust accurate method. However, this approach requires a higher number cameras considered optimization: practice, more than 30 are necessary, while typical vision-only BA can...

Constrained key-frame based local bundle adjustment is at the core of many recent systems that address problem large-scale, georeferenced SLAM on a monocular camera and data from inexpensive sensors and/or databases. The majority these methods, however, impose constraints result proprioceptive (e.g. IMUs, GPS, Odometry) while ignoring possibility explicitly constraining structure point cloud) resulting reconstruction process. Moreover, research on-line interactions between deep learning...

Model-based RL/control have gained significant traction in robotics. Yet, these approaches often remain data-inefficient and lack the explainability of hand-engineered solutions. This makes them difficult to debug/integrate safety-critical settings. However, many systems, prior knowledge environment kinematics/dynamics is available. Incorporating such priors can help address aforementioned problems by reducing problem complexity need for exploration, while also facilitating expression...

The Novelty Search (NS) algorithm was proposed more than a decade ago. However, the mechanisms behind its empirical success are still not well formalized/understood. This short note focuses on effects of archive exploration. Experimental evidence from few application domains suggests that archive-based NS performs in general better when is solely computed with respect to population. An argument often encountered literature prevents exploration backtracking or cycling, i.e. revisiting...

Quality-Diversity is a branch of stochastic optimization that often applied to problems from the Reinforcement Learning and control domains in order construct repertoires well-performing policies/skills exhibit diversity with respect behavior space. Such archives are usually composed finite number reactive agents which each associated unique descriptor, instantiating descriptors outside coarsely discretized space not straight-forward. While few recent works suggest solutions issue,...

Model-based reinforcement learning and control have demonstrated great potential in various sequential decision making problem domains, including robotics settings. However, real-world systems often present challenges that limit the applicability of those methods. In particular, we note two problems jointly happen many industrial systems: first, irregular/asynchronous observations actions and, second, dramatic changes environment dynamics from an episode to another (e.g <italic...

While the field of Quality-Diversity (QD) has grown into a distinct branch stochastic optimization, few problems, in particular locomotion and navigation tasks, have become de facto standards. Are such benchmarks sufficient? they representative key challenges faced by QD algorithms? Do provide ability to focus on one challenge properly disentangling it from others? much predictive power terms scalability generalization? Existing are not standardized, there is currently no MNIST equivalent...

Model-based Reinforcement Learning and Control have demonstrated great potential in various sequential decision making problem domains, including robotics settings. However, real-world systems often present challenges that limit the applicability of those methods. In particular, we note two problems jointly happen many industrial systems: 1) Irregular/asynchronous observations actions 2) Dramatic changes environment dynamics from an episode to another (e.g. varying payload inertial...