Reinforcement learning (RL) agents epitomize artificial intelligence by embodying adaptive prowess, navigating intricate knowledge landscapes through iterative trial and error, and dynamically assimilating environmental insights to autonomously evolve and optimize their decision-making capabilities. Developing generalist RL agents that can perform diverse tasks in complex environments is a challenging task that requires numerous reward functions. However,…

Large Language Models (LLMs)  have successfully utilized the power of Artificial Intelligence (AI) sub-fields, including Natural Language Processing (NLP), Natural Language Generation (NLG), and Computer Vision. With LLMs, the creation of vision-language models that can reason complexly about images, respond to queries pertaining to images, and describe images in natural language has been made possible.…

The progress in neural rendering has brought significant breakthroughs in reconstructing scenes and generating new viewpoints. However, its effectiveness largely depends on the precise pre-computation of camera poses. To minimize this problem, many efforts have been made to train Neural Radiance Fields (NeRFs) without precomputed camera poses. However, the implicit representation of NeRFs makes it…

Object perception in images and videos unleashes the power of machines to decipher the visual world. Like virtual sleuths, computer vision systems scour pixels, recognizing, tracking, and understanding the myriad objects that paint the canvas of digital experiences. This technological prowess, fueled by deep learning magic, opens doors to transformative applications – from self-driving cars…

Video super-resolution, aiming to elevate the quality of low-quality videos to high fidelity, faces the daunting challenge of addressing diverse and intricate degradations commonly found in real-world scenarios. Unlike previous focuses on synthetic or specific camera-related degradations, the complexity arises from multiple unknown factors like downsampling, noise, blur, flickering, and video compression. While recent CNN-based…

3D human motion reconstruction is a complex process that involves accurately capturing and modeling the movements of a human subject in three dimensions. This job becomes even more challenging when dealing with videos captured by a moving camera in real-world settings, as they often contain issues like foot sliding. However, a team of researchers from…

The field of pose estimation, which involves determining the position and orientation of an object in space, is a rapidly evolving area, with researchers continuously developing new methods to improve its accuracy and performance. Researchers from three highly regarded institutions – Tsinghua Shenzhen International Graduate School, Shanghai AI Laboratory, and Nanyang Technological University – have…

The Segment Anything Model (SAM) is an AI-powered model that segments images for object detection and recognition. It is an effective solution for various computer vision tasks. However, SAM is not optimized for edge devices, which can lead to retarded performance and high resource consumption. Researchers from S-Lab Nanyang Technological University and Shanghai Artificial Intelligence…

Researchers from Carnegie Mellon University and Google DeepMind have collaborated to develop RoboTool, a system leveraging Large Language Models (LLMs) to imbue robots with the ability to creatively use tools in tasks involving implicit physical constraints and long-term planning. The system comprises four key components:  Analyzer for interpreting natural language Planner for generating strategies Calculator…