A5075643825- Reinforcement Learning in Robotics
- Multimodal Machine Learning Applications
- Robot Manipulation and Learning
- Domain Adaptation and Few-Shot Learning
- Algal biology and biofuel production
- Marine and coastal ecosystems
- Advanced Neural Network Applications
- Aquatic Ecosystems and Phytoplankton Dynamics
- Robotic Path Planning Algorithms
- Adversarial Robustness in Machine Learning
- Human Pose and Action Recognition
- Advanced Image Processing Techniques
- Video Surveillance and Tracking Methods
- Microwave Engineering and Waveguides
- Microbial Metabolism and Applications
- Advanced machining processes and optimization
- Advanced Vision and Imaging
- 3D Printing in Biomedical Research
- Neonatal Health and Biochemistry
- Heme Oxygenase-1 and Carbon Monoxide
- Anomaly Detection Techniques and Applications
- Integrated Circuits and Semiconductor Failure Analysis
- Pharmacological Effects of Natural Compounds
- Methemoglobinemia and Tumor Lysis Syndrome
- Generative Adversarial Networks and Image Synthesis
University of California, Berkeley
2022-2024
Nanjing University of Science and Technology
2023-2024
Cornell University
2020-2024
Berkeley College
2024
Fujian Agriculture and Forestry University
2021-2022
First Affiliated Hospital of Jinan University
2022
Innovation Academy for Microsatellites of Chinese Academy of Sciences
2021
General Research Institute for Nonferrous Metals (China)
2021
Grinm Advanced Materials (China)
2021
Shanghai Micro Satellite Engineering Center
2021
Datasets drive vision progress, yet existing driving datasets are impoverished in terms of visual content and supported tasks to study multitask learning for autonomous driving. Researchers usually constrained a small set problems on one dataset, while real-world computer applications require performing various complexities. We construct BDD100K, the largest video dataset with 100K videos 10 evaluate exciting progress image recognition algorithms The possesses geographic, environmental,...
Building home assistant robots has long been a goal for vision and robotics researchers. To achieve this task, simulated environment with physically realistic simulation, sufficient articulated objects, transferability to the real robot is indispensable. Existing environments these requirements simulation different levels of simplification focus. We take one step further in constructing an that supports household tasks training learning algorithm. Our work, SAPIEN, physics-rich hosts...
Datasets drive vision progress, yet existing driving datasets are impoverished in terms of visual content and supported tasks to study multitask learning for autonomous driving. Researchers usually constrained a small set problems on one dataset, while real-world computer applications require performing various complexities. We construct BDD100K, the largest video dataset with 100K videos 10 evaluate exciting progress image recognition algorithms The possesses geographic, environmental,...
We propose an adversarial defense method that achieves state-of-the-art performance among attack-agnostic methods while also maintaining robustness to input resolution, scale of perturbation, and dataset size. Based on convolutional sparse coding, we construct a stratified low-dimensional quasi-natural image space faithfully approximates the natural removing perturbations. introduce novel Sparse Transformation Layer (STL) in between first layer neural network efficiently project images into...
A mechanistic understanding of algal growth is essential for maintaining a sustainable environment in an era climate change and population expansion. It known that tightly controlled by complex interactive physical chemical conditions. Many mathematical models have been proposed to describe the relation environmental parameters, but experimental verification has difficult due lack tools measure cell under precise As such, current depend on specific testing systems, fitted kinetic constants...
Simulation offers a promising approach for cheaply scaling training data generalist policies. To scalably generate from diverse and realistic tasks, existing algorithms either rely on large language models (LLMs) that may hallucinate tasks not interesting robotics; or digital twins, which require careful real-to-sim alignment are hard to scale. address these challenges, we introduce Video2Policy, novel framework leverages internet RGB videos reconstruct based everyday human behavior. Our...
Vision-Language-Action (VLA) models aim to predict robotic actions based on visual observations and language instructions. Existing approaches require fine-tuning pre-trained visionlanguage (VLMs) as features are independently fed into downstream policies, degrading the semantic alignments. We propose OTTER, a novel VLA architecture that leverages these existing alignments through explicit, text-aware feature extraction. Instead of processing all features, OTTER selectively extracts passes...
Many tasks in artificial intelligence require the collaboration of multiple agents. We exam deep reinforcement learning for multi-agent domains. Recent research efforts often take form two seemingly conflicting perspectives, decentralized perspective, where each agent is supposed to have its own controller; and centralized one assumes there a larger model controlling all In this regard, we revisit idea master-slave architecture by incorporating both perspectives within framework. Such...
Prodigiosin is a promising secondary metabolite produced mainly by Serratia strains. To study the global regulatory mechanism of prodigiosin biosynthesis, mutagenesis library containing 23,000 mutant clones was constructed with EZ-Tn5 transposon, and 114 in showed altered production ability. For 37 clones, transposon insertion occurred on biosynthetic cluster genes; inserted genes 77 belonged to 33 different outside genes. These can be divided into transcription-regulating genes, membrane...
Reinforcement learning has seen wide success in finetuning large language models to better align with instructions via human feedback. The so-called algorithm, Learning Human Feedback (RLHF) demonstrates impressive performance on the GPT series models. However, underlying (RL) algorithm is complex and requires an additional training pipeline for reward value networks. In this paper, we consider alternative approach: converting feedback instruction by relabeling original one model alignment a...
Open-vocabulary generalization requires robotic systems to perform tasks involving complex and diverse environments task goals. While the recent advances in vision language models (VLMs) present unprecedented opportunities solve unseen problems, how utilize their emergent capabilities control robots physical world remains an open question. In this paper, we MOKA (Marking Keypoint Affordances), approach that employs VLMs manipulation specified by free-form descriptions. At heart of our is a...
Overuse of acetaminophen (APAP) is a major cause drug-induced liver failure at the clinics. Apigenin (API) natural flavonoid derived from Matricaria chamomilla. The aim present study was to investigate amelioration function API in APAP-induced hepatotoxicity both vitro and vivo its potential mechanisms. Analysis results activities serum alanine aspartate aminotransferases (ALT AST), malondialdehyde, myeloperoxidase (MPO), reactive oxygen species (ROS) demonstrated therapeutic effects API....
Consider an imitation learning problem that the imitator and expert have different dynamics models. Most of current methods fail because they focus on imitating actions. We propose a novel state alignment-based method to train follow sequences in demonstrations as much possible. The alignment comes from both local global perspectives we combine them into reinforcement framework by regularized policy update objective. show superiority our standard settings where
Since 1960s, many countries led by the United States and Russia have done a lot of research on space nuclear reactor power. Moderator is an important functional component in operating with thermal neutron spectrum. This paper reviews moderation principle, main existing problems, status as well challenges ahead possible solutions moderator materials reactor. Aiming at problems hydrogen cracking, low utilization rate fuel loss materials, assumption using high entropy alloy hydride material put...
An agent that has well understood the environment should be able to apply its skills for any given goals, leading fundamental problem of learning Universal Value Function Approximator (UVFA). A UVFA learns predict cumulative rewards between all state-goal pairs. However, empirically, value function long-range goals is always hard estimate and may consequently result in failed policy. This presented challenges process capability neural networks. We propose a method address this issue large...
A general-purpose robot should be able to master a wide range of tasks and quickly learn novel one by leveraging past experiences. One-shot imitation learning (OSIL) approaches this goal training an agent with (pairs of) expert demonstrations, such that at test time, it can directly execute new task from just demonstration. However, so far framework has been limited on many variations task, testing other unseen but similar the same task. In work, we push for higher level generalization...
Visual model-based reinforcement learning (RL) has the potential to enable sample-efficient robot from visual observations. Yet current approaches typically train a single model end-to-end for both representations and dynamics, making it difficult accurately interaction between robots small objects. In this work, we introduce RL framework that decouples representation dynamics learning. Specifically, an autoencoder with convolutional layers vision transformers (ViT) reconstruct pixels given...
Video prediction is an important yet challenging problem; burdened with the tasks of generating future frames and learning environment dynamics. Recently, autoregressive latent video models have proved to be a powerful tool, by separating into two sub-problems: pre-training image generator model, followed model in space generator. However, successfully high-fidelity high-resolution videos has seen. In this work, we investigate how train capable predicting minimal modification existing...
Prodigiosin is a promising secondary metabolite mainly produced by Serratia marcescens . The production of prodigiosin S. regulated different kinds regulatory systems, including the EnvZ/OmpR system. In this study, we demonstrated that factor OmpR positively in FZSF02 directly binding to promoter region biosynthesis cluster with lacZ reporter assay and electrophoretic mobility shift (EMSA). sequence pig was identified DNase I footprinting assay. We further demonstrate regulates its own...
In recent years, the transformer architecture has become de facto standard for machine learning algorithms applied to natural language processing and computer vision. Despite notable evidence of successful deployment this in context robot learning, we claim that vanilla transformers do not fully exploit structure problem. Therefore, propose Body Transformer (BoT), an leverages embodiment by providing inductive bias guides process. We represent body as a graph sensors actuators, rely on...