Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
49 篇论文 · 按点赞排序
Yang Liu, Pengxiang Ding, Siteng Huang +3 authors
A new video-language model, PiTe, uses trajectory-guided pixel-temporal alignment to achieve superior performance across various multimodal video tasks by leveraging a large pre-training dataset with precise object trajectories.
Thomas Hanwen Zhu, Ruining Li, Tomas Jakab
DreamHOI uses text-to-image diffusion models and a dual implicit-explicit skinned mesh representation to synthesize realistic zero-shot human-object interactions.
Haritheja Etukuru, Norihito Naka, Zijin Hu +7 authors
Robot models, particularly those trained with large amounts of data, have recently shown a plethora of real-world manipulation and navigation capabilities. Several independent efforts have shown that given sufficient training data in an environment, robot policies can generalize to demonstrated variations in that environment. However, needing to finetune robot models to every new environment stands in stark contrast to models in language or vision that can be deployed zero-shot for open-world problems. In this work, we present Robot Utility Models (RUMs), a framework for training and deploying zero-shot robot policies that can directly generalize to new environments without any finetuning. To create RUMs efficiently, we develop new tools to quickly collect data for mobile manipulation tasks, integrate such data into a policy with multi-modal imitation learning, and deploy policies on-device on Hello Robot Stretch, a cheap commodity robot, with an external mLLM verifier for retrying. We train five such utility models for opening cabinet doors, opening drawers, picking up napkins, picking up paper bags, and reorienting fallen objects. Our system, on average, achieves 90% success rate in unseen, novel environments interacting with unseen objects. Moreover, the utility models can also succeed in different robot and camera set-ups with no further data, training, or fine-tuning. Primary among our lessons are the importance of training data over training algorithm and policy class, guidance about data scaling, necessity for diverse yet high-quality demonstrations, and a recipe for robot introspection and retrying to improve performance on individual environments. Our code, data, models, hardware designs, as well as our experiment and deployment videos are open sourced and can be found on our project website: https://robotutilitymodels.com
Tianhe Lu, Jizhan Fang, Yunzhi Yao +3 authors
A benchmark dataset (CKnowEdit) is introduced to address knowledge gaps in LLMs when generating Chinese content, highlighting areas requiring improvement in knowledge editing techniques.
Sacha Muller, António Loison, Bilel Omrani +1 authors
A retrieved-augmented generation (RAG) evaluation benchmark reveals limitations in existing automated frameworks, suggesting the need for unit test evaluations and fine-tuning LLMs on GPT-4's reasoning.
Dafei Qin, Hongyang Lin, Qixuan Zhang +7 authors
GauFace, a Gaussian Splatting representation, and TransGS, a diffusion transformer, enable efficient and high-fidelity rendering of facial assets with real-time performance across multiple platforms.
Qiuhong Shen, Xingyi Yang, Xinchao Wang
A globally optimal linear programming solver for 3D Gaussian Splatting segmentation is proposed, offering significant speed and robustness improvements.
Yan-Bo Lin, Yu Tian, Linjie Yang +2 authors
A generative video-music Transformer framework uses semantic and beat alignment to generate diverse and realistic background music from video inputs leveraging a large-scale dataset.
Mukhammadsaid Mamasaidov, Abror Shopulatov
The study provides datasets and neural models for Karakalpak language translation, demonstrating improvements over existing baselines.
Yuhao Wu, Ming Shan Hee, Zhiqing Hu +1 authors
A new benchmark, Spinning the Golden Thread, evaluates long-form text generation by assessing models' ability to incorporate specific events within generated text, revealing performance gaps compared to existing long-context tasks.
Tyler Bonnen, Stephanie Fu, Yutong Bai +5 authors
Human performance significantly outperforms common vision models in a zero-shot 3D shape inference task with varied viewpoints, as evaluated using a cognitive science-based benchmark.
Ben Bogin, Kejuan Yang, Shashank Gupta +5 authors
SUPER, a novel benchmark, evaluates Large Language Models' ability to autonomously set up and execute tasks from research repositories, highlighting significant challenges in this domain.
Fei Ye, Zaixiang Zheng, Dongyu Xue +7 authors
ProteinBench is an evaluation framework that enhances transparency in protein foundation models through taxonomic classification and multi-metric assessments.
Hanyu Jiang, Jian Xue, Xing Lan +2 authors
MVLLaVA combines multi-view diffusion models with LLaVA to efficiently generate novel views based on user instructions for diverse tasks.
Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin +2 authors
A 3D object detection model trained on a combined indoor dataset achieves superior performance across multiple benchmarks using a unified label space and transformer encoder architecture.
Sherry Yang, Simon Batzner, Ruiqi Gao +7 authors
GenMS combines a language model, diffusion model, and graph neural network to generate crystal structures from natural language, improving satisfaction of user requests and energy efficiency.
Yi Cui
Evaluation of large language models on the WebApp1K benchmark indicates varied performance in generating correct web application code, highlighting the need for improvements in model reliability.
Jiahui Liu, Xin Wen, Shizhen Zhao +2 authors
SyncOOD uses generative models to create synthetic OOD samples, enhancing OOD object detection by optimally setting ID/OOD decision boundaries.
Archana Swaminathan, Anubhav Gupta, Kamal Gupta +3 authors
LEIA learns view-invariant latent representations for dynamic 3D objects using hypernetworks to model different states and generate novel articulations.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号