A5042425097- Reinforcement Learning in Robotics
- Autonomous Vehicle Technology and Safety
- Evolutionary Algorithms and Applications
- Robotic Locomotion and Control
- Video Surveillance and Tracking Methods
- Human Pose and Action Recognition
- Evolutionary Game Theory and Cooperation
- Robotic Path Planning Algorithms
- Evolution and Genetic Dynamics
- Adaptive Dynamic Programming Control
- Machine Learning and Algorithms
- Forecasting Techniques and Applications
- Time Series Analysis and Forecasting
- Advanced Bandit Algorithms Research
- Stock Market Forecasting Methods
- Domain Adaptation and Few-Shot Learning
- Data Stream Mining Techniques
- Metaheuristic Optimization Algorithms Research
- Natural Language Processing Techniques
- Twentieth Century Scientific Developments
- Advanced Control Systems Optimization
- Advanced Multi-Objective Optimization Algorithms
- Ethics and Social Impacts of AI
- Control Systems and Identification
- Anomaly Detection Techniques and Applications
Sorbonne Université
2019-2023
Centre National de la Recherche Scientifique
2019-2023
Institut Systèmes Intelligents et de Robotique
2019-2023
SoftBank Robotics (France)
2020-2023
Huawei Technologies (France)
2023
Université Sorbonne Nouvelle
2021
ETH Zurich
2017-2018
We present a learning-based mapless motion planner by taking the sparse 10-dimensional range findings and target position with respect to mobile robot coordinate frame as input continuous steering commands output. Traditional planners for ground robots laser sensor mostly depend on obstacle map of navigation environment where both highly precise building work are indispensable. show that, through an asynchronous deep reinforcement learning method, can be trained end-to-end without any...
This paper reports on a data-driven, interaction-aware motion prediction approach for pedestrians in environments cluttered with static obstacles. When navigating such workspaces shared humans, robots need accurate predictions of the surrounding pedestrians. Human navigation behavior is mostly influenced by their and obstacles vicinity. In this we introduce new model based Long-Short Term Memory (LSTM) neural networks, which able to learn human from demonstrated data. To best our knowledge,...
We present a learning-based mapless motion planner by taking the sparse 10-dimensional range findings and target position with respect to mobile robot coordinate frame as input continuous steering commands output. Traditional planners for ground robots laser sensor mostly depend on obstacle map of navigation environment where both highly precise building work are indispensable. show that, through an asynchronous deep reinforcement learning method, can be trained end-to-end without any...
Pre-trained foundation models (FMs) have shown exceptional performance in univariate time series forecasting tasks. However, several practical challenges persist, including managing intricate dependencies among features and quantifying uncertainty predictions. This study aims to tackle these critical limitations by introducing adapters; feature-space transformations that facilitate the effective use of pre-trained FMs for multivariate Adapters operate projecting inputs into a suitable latent...
Performing Reinforcement Learning in sparse rewards settings, with very little prior knowledge, is a challenging problem since there no signal to properly guide the learning process. In such situations, good search strategy fundamental. At same time, not having adapt algorithm every single desirable. Here we introduce TAXONS, Task Agnostic eXploration of Outcome spaces through Novelty and Surprise algorithm. Based on population-based divergent-search approach, it learns set diverse policies...
Evolvability is an important feature that impacts the ability of evolutionary processes to find interesting novel solutions and deal with changing conditions problem solve. The estimation evolvability not straight-forward generally too expensive be directly used as selective pressure in process. Indirectly promoting a side effect other easier faster compute selection pressures would thus advantageous. In unbounded behavior space, it has already been shown evolvable individuals naturally...
Reward-based optimization algorithms require both exploration, to find rewards, and exploitation, maximize performance. The need for efficient exploration is even more significant in sparse reward settings, which performance feedback given sparingly, thus rendering it unsuitable guiding the search process. In this work, we introduce SparsE Reward Exploration via Novelty Emitters (SERENE) algorithm, capable of efficiently exploring a space, as well optimizing rewards found potentially...
Learning optimal policies in sparse rewards settings is difficult as the learning agent has little to no feedback on quality of its actions. In these situations, a good strategy focus exploration, hopefully leading discovery reward signal improve on. A algorithm capable dealing with this kind setting be able (1) explore possible behaviors and (2) exploit any discovered reward. Exploration algorithms have been proposed that require definition low-dimension behavior space, which generated by...
In this paper we focus on developing a control algorithm for multi-terrain tracked robots with flippers using reinforcement learning (RL) approach. The work is based the deep deterministic policy gradient (DDPG) algorithm, proven to be very successful in simple simulation environments. works an end-to-end fashion order continuous position of flippers. This approach makes it easy apply controller wide array circumstances, but huge flexibility comes cost increased difficulty solution....
In model-based reinforcement learning, most algorithms rely on simulating trajectories from one-step models of the dynamics learned data. A critical challenge this approach is compounding prediction errors as length trajectory grows. paper we tackle issue by using a multi-step objective to train models. Our weighted sum mean squared error (MSE) loss at various future horizons. We find that new particularly useful when data noisy (additive Gaussian noise in observations), which often case...
We propose Embodied AI as the next fundamental step in pursuit of Artificial General Intelligence, juxtaposing it against current advancements, particularly Large Language Models. traverse evolution embodiment concept across diverse fields - philosophy, psychology, neuroscience, and robotics to highlight how EAI distinguishes itself from classical paradigm static learning. By broadening scope AI, we introduce a theoretical framework based on cognitive architectures, emphasizing perception,...
Transformer-based architectures achieved breakthrough performance in natural language processing and computer vision, yet they remain inferior to simpler linear baselines multivariate long-term forecasting. To better understand this phenomenon, we start by studying a toy forecasting problem for which show that transformers are incapable of converging their true solution despite high expressive power. We further identify the attention as being responsible low generalization capacity. Building...
The emerging zero-shot capabilities of Large Language Models (LLMs) have led to their applications in areas extending well beyond natural language processing tasks. In reinforcement learning, while LLMs been extensively used text-based environments, integration with continuous state spaces remains understudied. this paper, we investigate how pre-trained can be leveraged predict context the dynamics Markov decision processes. We identify handling multivariate data and incorporating control...
We introduce Agent K v1.0, an end-to-end autonomous data science agent designed to automate, optimise, and generalise across diverse tasks. Fully automated, v1.0 manages the entire life cycle by learning from experience. It leverages a highly flexible structured reasoning framework enable it dynamically process memory in nested structure, effectively accumulated experience stored handle complex optimises long- short-term selectively storing retrieving key information, guiding future...
Evolvability is an important feature that impacts the ability of evolutionary processes to find interesting novel solutions and deal with changing conditions problem solve. The estimation evolvability not straightforward generally too expensive be directly used as selective pressure in process. Indirectly promoting a side effect other easier faster compute selection pressures would thus advantageous. In unbounded behavior space, it has already been shown evolvable individuals naturally...
No abstract available.
In model-based reinforcement learning (MBRL), most algorithms rely on simulating trajectories from one-step dynamics models learned data. A critical challenge of this approach is the compounding prediction errors as length trajectory grows. paper we tackle issue by using a multi-timestep objective to train models. Our weighted sum loss function (e.g., negative log-likelihood) at various future horizons. We explore and test range weights profiles. find that exponentially decaying lead...
This paper reports on a data-driven, interaction-aware motion prediction approach for pedestrians in environments cluttered with static obstacles. When navigating such workspaces shared humans, robots need accurate predictions of the surrounding pedestrians. Human navigation behavior is mostly influenced by their and obstacles vicinity. In this we introduce new model based Long-Short Term Memory (LSTM) neural networks, which able to learn human from demonstrated data. To best our knowledge,...
Performing Reinforcement Learning in sparse rewards settings, with very little prior knowledge, is a challenging problem since there no signal to properly guide the learning process. In such situations, good search strategy fundamental. At same time, not having adapt algorithm every single desirable. Here we introduce TAXONS, Task Agnostic eXploration of Outcome spaces through Novelty and Surprise algorithm. Based on population-based divergent-search approach, it learns set diverse policies...
In the Reinforcement Learning (RL) framework, learning is guided through a reward signal. This means that in situations of sparse rewards agent has to focus on exploration, order discover which action, or set actions leads reward. RL agents usually struggle with this. Exploration Quality-Diversity (QD) methods. this thesis, we approach problem these algorithms, and particular Novelty Search (NS). method only focuses diversity possible policies behaviors. The first part thesis representation...
In the last decade, reinforcement learning successfully solved complex control tasks and decision-making problems, like Go board game. Yet, there are few success stories when it comes to deploying those algorithms real-world scenarios. One of reasons is lack guarantees dealing with avoiding unsafe states, a fundamental requirement in critical engineering systems. this paper, we introduce Guided Safe Shooting (GuSS), model-based RL approach that can learn systems minimal violations safety...