A5101632944- Multimodal Machine Learning Applications
- Robotic Path Planning Algorithms
- Robotics and Sensor-Based Localization
- Reinforcement Learning in Robotics
- Advanced Neural Network Applications
- Evolutionary Algorithms and Applications
- Natural Language Processing Techniques
- Domain Adaptation and Few-Shot Learning
- Robot Manipulation and Learning
- Evacuation and Crowd Dynamics
- Human Pose and Action Recognition
- Autonomous Vehicle Technology and Safety
- Neural dynamics and brain function
Southeast University
2024
Southeast University
2023
Multi-task learning is an important problem in reinforcement learning. Training multiple tasks together brings benefits from the shared useful information across different and often achieves higher performance compared to single-task However, it remains unclear how parameters network should be reused tasks. Instead of naively sharing all tasks, we propose attention-based mixture experts multi-task approach learn a compositional policy for each task. The expert networks task-specific skills...
In this paper, we present XuanCe, a comprehensive and unified deep reinforcement learning (DRL) library designed to be compatible with PyTorch, TensorFlow, MindSpore. XuanCe offers wide range of functionalities, including over 40 classical DRL multi-agent algorithms, the flexibility easily incorporate new algorithms environments. It is versatile that supports CPU, GPU, Ascend, can executed on various operating systems such as Ubuntu, Windows, MacOS, EulerOS. Extensive benchmarks conducted...
Zero-shot object navigation is a challenging task for home-assistance robots. This emphasizes visual grounding, commonsense inference and locomotion abilities, where the first two are inherent in foundation models. But part, most works still depend on map-based planning approaches. The gap between RGB space map makes it difficult to directly transfer knowledge from models tasks. In this work, we propose Pixel-guided Navigation skill (PixNav), which bridges embodied task. It straightforward...
While large language models (LLMs) are successful in completing various processing tasks, they easily fail to interact with the physical world by generating control sequences properly. We find that main reason is LLMs not grounded world. Existing LLM-based approaches circumvent this problem relying on additional pre-defined skills or pre-trained sub-policies, making it hard adapt new tasks. In contrast, we aim address and explore possibility prompt accomplish a series of robotic manipulation...
Recently, learning-based approaches show promising results in navigation tasks. However, the poor generalization capability and simulation-reality gap prevent a wide range of applications. We consider problem improving mobile robots achieving sim-to-real transfer for skills. To that end, we propose cross-modal fusion method knowledge framework better generalization. This is realized by teacher-student distillation architecture. The teacher learns discriminative representation near-perfect...
Recently, learning-based approaches show promising results in navigation tasks. However, the poor generalization capability and simulation-reality gap prevent a wide range of applications. We consider problem improving mobile robots achieving sim-to-real transfer for skills. To that end, we propose cross-modal fusion method knowledge framework better generalization. This is realized by teacher-student distillation architecture. The teacher learns discriminative representation near-perfect...
In recent years, learning-based approaches have demonstrated significant promise in addressing intricate navigation tasks. Traditional methods for training deep neural network policies rely on meticulously designed reward functions or extensive teleoperation datasets as demonstrations. However, the former is often confined to simulated environments, and latter demands substantial human labor, making it a time-consuming process. Our vision robots autonomously learn skills adapt their...