A5071943173- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Adversarial Robustness in Machine Learning
- Machine Learning and Algorithms
- Robotics and Sensor-Based Localization
- Robotic Path Planning Algorithms
- Robot Manipulation and Learning
- Reinforcement Learning in Robotics
- Anomaly Detection Techniques and Applications
- Human Motion and Animation
- Speech and dialogue systems
- Human Pose and Action Recognition
- Advanced Image and Video Retrieval Techniques
- Topic Modeling
- Cognitive Science and Education Research
- Epistemology, Ethics, and Metaphysics
- Psychological and Educational Research Studies
- Data Stream Mining Techniques
- Advanced Data Processing Techniques
- Generative Adversarial Networks and Image Synthesis
- Robotic Locomotion and Control
- Software Testing and Debugging Techniques
- Underwater Vehicles and Communication Systems
Boston Dynamics (United States)
2024
University of Pennsylvania
2020-2024
University of Michigan–Ann Arbor
2024
Robot learning has emerged as a promising tool for taming the complexity and diversity of real world. Methods based on high-capacity models, such deep networks, hold promise providing effective generalization to wide range open-world environments. However, these same methods typically require large amounts diverse training data generalize effectively. In contrast, most robotic experiments are small-scale, single-domain, single-robot. This leads frequent tension in learning: how can we learn...
Robot learning holds the promise of policies that generalize broadly. However, such generalization requires sufficiently diverse datasets task interest, which can be prohibitively expensive to collect. In other fields, as computer vision, it is common utilize shared, reusable datasets, ImageNet, overcome this challenge, but has proven difficult in robotics. paper, we ask: what would take enable practical data reuse robotics for end-to-end skill learning? We hypothesize key use with multiple...
We consider the problems of exploration and pointgoal navigation in previously unseen environments, where spatial complexity indoor scenes partial observability constitute these tasks challenging. argue that learning occupancy priors over maps provides significant advantages towards addressing problems. To this end, we present a novel planning framework first learns to generate beyond field-of-view agent, second leverages model uncertainty generated areas formulate path selection policies...
Traversing terrain with good traction is crucial for achieving fast off-road navigation. Instead of manually designing costs based on features, existing methods learn properties directly from data via self-supervision to automatically penalize trajectories moving through undesirable terrain, but challenges remain in properly quantifying and mitigating the risk due uncertainty learned models. To this end, we present evidential autonomy (EVORA), a unified framework uncertainty-aware model plan...
Large-scale robotic policies trained on data from diverse tasks and platforms hold great promise for enabling general-purpose robots; however, reliable generalization to new environment conditions remains a major challenge. Toward addressing this challenge, we propose novel approach uncertainty-aware deployment of pre-trained language-conditioned imitation learning agents. Specifically, use temperature scaling calibrate these models exploit the calibrated model make decisions by aggregating...
We present a fully autonomous real-world RL framework for mobile manipulation that can learn policies without extensive instrumentation or human supervision. This is enabled by 1) task-relevant autonomy, which guides exploration towards object interactions and prevents stagnation near goal states, 2) efficient policy learning leveraging basic task knowledge in behavior priors, 3) formulating generic rewards combine human-interpretable semantic information with low-level, fine-grained...
Good pre-trained visual representations could enable robots to learn visuomotor policy efficiently. Still, existing take a one-size-fits-all-tasks approach that comes with two important drawbacks: (1) Being completely task-agnostic, these cannot effectively ignore any task-irrelevant information in the scene, and (2) They often lack representational capacity handle unconstrained/complex real-world scenes. Instead, we propose train large combinatorial family of organized by scene entities:...
Landmark-based navigation (e.g. go to the wooden desk) and relative positional move 5 meters forward) are distinct challenges solved very differently in existing robotics methodology. We present a new dataset, OC-VLN, order distinctly evaluate grounding object-centric natural language instructions method for performing landmark-based navigation. also propose Natural Language grounded SLAM (NL-SLAM), ground instruction robot observations poses. actively perform NL-SLAM follow instructions....
Collecting new experience is costly in many robotic tasks, so determining how to efficiently explore a environment learn as much possible few trials an important problem for robotics. In this paper, we propose method exploring the purpose of learning dynamics model. Our key idea minimize score given by discriminator network objective planner which chooses actions. This optimized jointly with prediction model and enables our active approach sample sequences observations actions result...
Style analysis of artwork in computer vision predominantly focuses on achieving results target image generation through optimizing understanding low level style characteristics such as brush strokes. However, fundamentally different techniques are required to computationally understand and control qualities art which incorporate higher characteristics. We study representations learned by neural network architectures incorporating these find variation features from triplets annotated...
Traversing terrain with good traction is crucial for achieving fast off-road navigation. Instead of manually designing costs based on features, existing methods learn properties directly from data via self-supervision to automatically penalize trajectories moving through undesirable terrain, but challenges remain properly quantify and mitigate the risk due uncertainty in learned models. To this end, work proposes a unified framework uncertainty-aware model plan risk-aware trajectories. For...
Understanding how humans leverage semantic knowledge to navigate unfamiliar environments and decide where explore next is pivotal for developing robots capable of human-like search behaviors. We introduce a zero-shot navigation approach, Vision-Language Frontier Maps (VLFM), which inspired by human reasoning designed towards unseen objects in novel environments. VLFM builds occupancy maps from depth observations identify frontiers, leverages RGB pre-trained vision-language model generate...