NEW
26년도 연구참여 학생 모집
POSTECH 컴퓨터비전연구실(CVLab)에서 2026년 하반기 연구참여 학생을 모집합니다. 현재 연구실에서 진행 중인 연구들과 관련하여 준비된 세부 연구 주제들은 다음과 같습니다.
3D Symmetry Detection under Partial Observation (지도: 조민수 교수)
Robust Robot Action Policies (지도: 곽수하 교수)
Multimodal Memory for AI Agent (지도: 곽수하 교수)
Adversarial Attacks for Embodied and Physical AI (지도: 손진희 교수)
Dexterous Grasp Generation (지도: 조민수 교수)
Active Perception for Object Manipulation (지도: 조민수 교수)
Robotic Manipulation in Cluttered Environments (지도: 조민수 교수)
각 주제에 대한 상세 내용은 아래에서 확인할 수 있습니다.
위 주제 외에도 컴퓨터비전, 로보틱스, 멀티모달 AI 전반에 걸쳐 다양한 연구가 진행되고 있으니, 나열된 주제와 정확히 일치하지 않더라도 관심 있는 학생은 언제든 원하는 교수님께 편하게 연락 주기 바랍니다!
1. 3D Symmetry Detection under Partial Observation
Goal: develop a 3D symmetry detection method robust to partial observations (occlusion, incomplete scans) and shape deformation (physical damage, noisy scans), then validate it on real 3D shapes captured with an RGBD camera. Building on the lab's work on correspondence-based symmetry detection, the intern will treat symmetry detection as a self-matching problem, using geometry-aware biases such as group-equivariance to handle noisy, incomplete real-world geometry.
Tasks: design/implement the pipeline, train and evaluate on synthetic and real RGBD data, and analyze robustness to occlusion and deformation.
Requirements: systematic programming skills and linear algebra, and basic knowledge of (or interest in) abstract algebra and group theory. Prior 3D vision/deep learning experience is a plus, not required.
Interested students: reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr)
Related Papers:
Wongyun Yu, Ahyun Seo, Minsu Cho. Axis-level Symmetry Detection with Group-equivariant Representation. ICCV, 2025
Marcelo Cicconet, David Hildebrand, Harry Elliott. Finding Mirror Symmetry via Registration and Optimal Symmetric Pairwise Assignment. ICCV Workshops, 2017
Niloy J. Mitra, Leonidas Guibas, Mark Pauly. Partial and Approximate Symmetry Detection for 3D Geometry. ACM SIGGRAPH, 2006
Related Research Page: Symmetry Analysis & Equivariant Learning
2. Robust Robot Action Policies
Goal: Enhance the robustness of robot action policies, such as vision-language-action (VLA) models and world-action models (WAM). Specifically, we explore neural network architectures and training strategies that enable these models to operate robustly against visual input degradation, paraphrasing of natural language commands, and the discrepancy between intended and executed actions. While we investigate individual degradation factors, our ultimate goal is to construct a single, unified model that remains robust across these various types of degradation.
Tasks:
Construct a benchmark to evaluate the robustness of action policy models under realistic degradation scenarios.
Develop solutions tailored to specific types of degradation.
Build a model capable of handling multiple types of degradation.
Conduct experiments using real robot hardware.
Requirements:
Proficient programming skills
Experience in computer vision and deep learning (preferred)
Interested students: Reach out with a short note on your background and interests via email to Prof. Suha Kwak (suha.kwak@postech.ac.kr).
Related papers:
Guo et al., On Robustness of Vision-language-action Model against Multi-modal Perturbations. In Proc. International Conference on Learning Representations (ICLR), 2026
Lee et al., Robust Promptable Video Object Segmentation, In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
Lee et al., GaRA-SAM: Robustifying Segment Anything Model with Gated-Rank Adaptation, In Proc. Conference on Neural Information Processing (NeurIPS), 2025
3. Multimodal Memory for AI Agent
Goal: Explore effective memory structures for AI agents, with a primary focus on knowledge graphs. By advancing the text-based knowledge graph construction and utilization techniques developed in our lab, we aim to build a system that automatically constructs and leverages knowledge graphs from multimodal data and AI agents that better utilize such memory structures. This system will then be applied to complex, multi-hop multimodal question-answering and life assistant AI systems.
Tasks:
Develop techniques that construct multimodal knowledge graph autonomously by detecting components within multimodal data and establishing their corresponding relationships.
Build an agent capable of exploring and navigating knowledge graphs effectively.
Apply these systems to question-answering and/or life assistant AI.
Requirements:
Proficient programming skills
Knowledge of graph-related algorithms (preferred)
Experience in deep learning (preferred)
Interested students: Reach out with a short note on your background and interests via email to Prof. Suha Kwak (suha.kwak@postech.ac.kr).
Related papers:
X. Zhu, Y. Xie, Y. Liu, Y. Li, and W. Hu. Knowledge graph-guided retrieval augmented generation. In Proc. North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), 2025
C. Bu, G. Chang, Z. Chen, C. Dang, Z. Wu, Y. He, and X. Wu. Query-driven multimodal GraphRAG: Dynamic local knowledge graph construction for online reasoning. In Findings of the association for computational linguistics (ACL), 2025
J. Yang et al. Egolife: Towards Egocentric Life Assistant. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
4. Adversarial Attacks for Embodied and Physical AI
Goal: Develop embodied and Physical AI systems that can perceive, reason, plan, and act reliably under adversarial attacks and unexpected real-world conditions. The project considers a broad range of systems, including embodied navigation agents, robot manipulation policies, world-model-based agents, and vision-language-action models. We investigate how attacks on sensors, visual observations, language instructions, internal world representations, and action execution propagate through the closed-loop perception–action pipeline.
Tasks:
Construct a benchmark for evaluating adversarial robustness across embodied navigation, robot manipulation, and language-guided control tasks.
Develop attacks targeting different stages of the embodied AI pipeline, including adversarial patches, sensor spoofing, visual prompt injection, corrupted world states, and action-level perturbations.
Analyze how local errors in perception or reasoning propagate into planning failures, incorrect trajectories, and unsafe physical actions.
Requirements:
Proficient programming skills
Basic knowledge of computer vision, deep learning, and linear algebra
Interest in adversarial robustness, embodied AI, or robot learning
Interested students: Reach out with a short note on your background and research interests via email to Prof. Jeany Son (jeany@postech.ac.kr).
Related papers:
J. Guo et al. RobustVLA: On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations. In Proc. International Conference on Learning Representations (ICLR), 2026.
L. Shen, M. Cui, and X. Yang. BadWorld: Adversarial Attacks on World Models. Preprint, 2026.
J. Li et al. AttackVLA: Benchmarking Adversarial and Backdoor Attacks on Vision-Language-Action Models. Preprint, 2025.
5. Dexterous Grasp Generation
Goal: Enable multi-fingered robotic hands to decide how to grasp an object. A grasp that is merely stable is often not the right one. Gripping a knife by the blade holds it firmly but makes it unusable, a grasp designed for a five-fingered hand may be infeasible on a four-fingered one. We therefore study generative models that produce robot grasps conditioned on what the object will be used for (task intent) and which hand is doing the grasping (hand morphology). While we investigate these factors individually, our ultimate goal is to build a single model that handles both together and works on real robot hardware.
Tasks:
Develop task-oriented grasp generation grounded in language or task specifications.
Design morphology-aware representations that generalize across different hand embodiments.
Build a unified model capable of handling both conditioning factors jointly.
Conduct experiments using real robot hardware.
Requirements:
Proficient programming skills
Experience in computer vision and deep learning (preferred)
Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr)
Related papers:
Lee et al., DextER: Language-driven Dexterous Grasp Generation with Embodied Reasoning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
Zhong et al., DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
Chen et al., Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy. In Proc. Robotics: Science and Systems (RSS), 2025
6. Active Perception for Object Manipulation
Goal: This project studies how a robot can actively move its wrist-mounted camera to a better viewpoint before grasping, rather than relying on a single fixed view. Most existing grasping pipelines assume a single, often fixed RGB-D observation, which frequently leaves target objects partially occluded, leading to inaccurate shape estimation and unstable grasps. From this partial observation, the system reconstructs the scene, selects the next-best camera pose expected to reveal previously occluded regions, moves the arm to acquire this view, and executes a dexterous grasp with a multi-fingered hand using the updated reconstruction. We aim to show that this active, viewpoint-aware strategy improves grasp success rate and robustness compared to single-view baselines, especially for objects with complex geometry and in cluttered scenes.
Tasks:
Build a simulation environment and validate grasp execution using target grasp poses.
Design and implement the perception–action closed loop (reconstruction, next-best-view selection, camera move, grasp).
Train and improve the next-best-view selection policy, and analyze failure cases.
Extend to real robot environments.
Requirements:
Proficient programming skills
Experience in deep learning
Experience with robot simulators (preferred)
Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr)
Related papers:
Wang, Kuancheng, et al. "Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation." arXiv preprint arXiv:2606.23625 (2026).
Park, Chunghyun, Seunghyeon Lee, and Minsu Cho. "Affostruction: 3D Affordance Grounding with Generative Reconstruction." arXiv preprint arXiv:2601.09211 (2026).
Lee, Junha, Eunha Park, and Minsu Cho. "Dexter: Language-driven dexterous grasp generation with embodied reasoning." arXiv preprint arXiv:2601.16046 (2026).
7. Robotic Manipulation in Cluttered Environments
Goal: Improve vision-language-action (VLA) models and other robot manipulation policies to operate reliably in realistic, obstacle-rich environments. Existing benchmarks mostly cover simple pick-and-place scenarios; in real-world settings, however, obstacle placements can make the target object inaccessible without interacting with surrounding objects, requiring the robot to reason about such interactions. We aim to build a simulation-based benchmark that evaluates robot policies in such environments across different difficulty levels and use it to analyze failures and improve performance.
Tasks:
Construct a benchmark for robot manipulation in realistic cluttered environments.
Evaluate VLA policies (e.g., OpenVLA, π0.5, GR00T N1.6) and analyze their failures.
Build training data and improve VLA performance through fine-tuning.
Extend to real robot environments.
Requirements:
Proficient programming skills
Experience in deep learning
Experience with robot simulators (preferred)
Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr).
Related Papers:
Physical Intelligence, π0.5: A Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054, 2025
Han et al., FetchBench: A Simulation Benchmark for Robot Fetching. In Proc. Conference on Robot Learning (CoRL), 2024
Danielczuk et al., Mechanical Search: Multi-Step Retrieval of a Target Object Occluded by Clutter. In Proc. IEEE International Conference on Robotics and Automation (ICRA), 2019
26년도 연구참여 학생 모집POSTECH 컴퓨터비전연구실(CVLab)에서 2026년 하반기 연구참여 학생을 모집합니다. 현재 연구실에서 진행 중인 연구들과 관련하여 준비된 세부 연구 주제들은 다음과 같습니다. 3D Symmetry Detection under Partial Observation (지도: 조민수 교수) Robust Robot Action Policies (지도: 곽수하 교수) Multimodal Memory for AI Agent (지도: 곽수하 교수) Adversarial Attacks for Embodied and Physical AI (지도: 손진희 교수) Dexterous Grasp Generation (지도: 조민수 교수) Active Perception for Object Manipulation (지도: 조민수 교수) Robotic Manipulation in Cluttered Environments (지도: 조민수 교수) 각 주제에 대한 상세 내용은 아래에서 확인할 수 있습니다. 위 주제 외에도 컴퓨터비전, 로보틱스, 멀티모달 AI 전반에 걸쳐 다양한 연구가 진행되고 있으니, 나열된 주제와 정확히 일치하지 않더라도 관심 있는 학생은 언제든 원하는 교수님께 편하게 연락 주기 바랍니다! 1. 3D Symmetry Detection under Partial Observation Goal: develop a 3D symmetry detection method robust to partial observations (occlusion, incomplete scans) and shape deformation (physical damage, noisy scans), then validate it on real 3D shapes captured with an RGBD camera. Building on the lab's work on correspondence-based symmetry detection, the intern will treat symmetry detection as a self-matching problem, using geometry-aware biases such as group-equivariance to handle noisy, incomplete real-world geometry. Tasks: design/implement the pipeline, train and evaluate on synthetic and real RGBD data, and analyze robustness to occlusion and deformation. Requirements: systematic programming skills and linear algebra, and basic knowledge of (or interest in) abstract algebra and group theory. Prior 3D vision/deep learning experience is a plus, not required. Interested students: reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr) Related Papers: Wongyun Yu, Ahyun Seo, Minsu Cho. Axis-level Symmetry Detection with Group-equivariant Representation. ICCV, 2025 Marcelo Cicconet, David Hildebrand, Harry Elliott. Finding Mirror Symmetry via Registration and Optimal Symmetric Pairwise Assignment. ICCV Workshops, 2017 Niloy J. Mitra, Leonidas Guibas, Mark Pauly. Partial and Approximate Symmetry Detection for 3D Geometry. ACM SIGGRAPH, 2006 Related Research Page: Symmetry Analysis & Equivariant Learning 2. Robust Robot Action Policies Goal: Enhance the robustness of robot action policies, such as vision-language-action (VLA) models and world-action models (WAM). Specifically, we explore neural network architectures and training strategies that enable these models to operate robustly against visual input degradation, paraphrasing of natural language commands, and the discrepancy between intended and executed actions. While we investigate individual degradation factors, our ultimate goal is to construct a single, unified model that remains robust across these various types of degradation. Tasks: Construct a benchmark to evaluate the robustness of action policy models under realistic degradation scenarios. Develop solutions tailored to specific types of degradation. Build a model capable of handling multiple types of degradation. Conduct experiments using real robot hardware. Requirements: Proficient programming skills Experience in computer vision and deep learning (preferred) Interested students: Reach out with a short note on your background and interests via email to Prof. Suha Kwak (suha.kwak@postech.ac.kr). Related papers: Guo et al., On Robustness of Vision-language-action Model against Multi-modal Perturbations. In Proc. International Conference on Learning Representations (ICLR), 2026 Lee et al., Robust Promptable Video Object Segmentation, In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026 Lee et al., GaRA-SAM: Robustifying Segment Anything Model with Gated-Rank Adaptation, In Proc. Conference on Neural Information Processing (NeurIPS), 2025 3. Multimodal Memory for AI Agent Goal: Explore effective memory structures for AI agents, with a primary focus on knowledge graphs. By advancing the text-based knowledge graph construction and utilization techniques developed in our lab, we aim to build a system that automatically constructs and leverages knowledge graphs from multimodal data and AI agents that better utilize such memory structures. This system will then be applied to complex, multi-hop multimodal question-answering and life assistant AI systems. Tasks: Develop techniques that construct multimodal knowledge graph autonomously by detecting components within multimodal data and establishing their corresponding relationships. Build an agent capable of exploring and navigating knowledge graphs effectively. Apply these systems to question-answering and/or life assistant AI. Requirements: Proficient programming skills Knowledge of graph-related algorithms (preferred) Experience in deep learning (preferred) Interested students: Reach out with a short note on your background and interests via email to Prof. Suha Kwak (suha.kwak@postech.ac.kr). Related papers: X. Zhu, Y. Xie, Y. Liu, Y. Li, and W. Hu. Knowledge graph-guided retrieval augmented generation. In Proc. North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), 2025 C. Bu, G. Chang, Z. Chen, C. Dang, Z. Wu, Y. He, and X. Wu. Query-driven multimodal GraphRAG: Dynamic local knowledge graph construction for online reasoning. In Findings of the association for computational linguistics (ACL), 2025 J. Yang et al. Egolife: Towards Egocentric Life Assistant. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025 4. Adversarial Attacks for Embodied and Physical AI Goal: Develop embodied and Physical AI systems that can perceive, reason, plan, and act reliably under adversarial attacks and unexpected real-world conditions. The project considers a broad range of systems, including embodied navigation agents, robot manipulation policies, world-model-based agents, and vision-language-action models. We investigate how attacks on sensors, visual observations, language instructions, internal world representations, and action execution propagate through the closed-loop perception–action pipeline. Tasks: Construct a benchmark for evaluating adversarial robustness across embodied navigation, robot manipulation, and language-guided control tasks. Develop attacks targeting different stages of the embodied AI pipeline, including adversarial patches, sensor spoofing, visual prompt injection, corrupted world states, and action-level perturbations. Analyze how local errors in perception or reasoning propagate into planning failures, incorrect trajectories, and unsafe physical actions. Requirements: Proficient programming skills Basic knowledge of computer vision, deep learning, and linear algebra Interest in adversarial robustness, embodied AI, or robot learning Interested students: Reach out with a short note on your background and research interests via email to Prof. Jeany Son (jeany@postech.ac.kr). Related papers: J. Guo et al. RobustVLA: On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations. In Proc. International Conference on Learning Representations (ICLR), 2026. L. Shen, M. Cui, and X. Yang. BadWorld: Adversarial Attacks on World Models. Preprint, 2026. J. Li et al. AttackVLA: Benchmarking Adversarial and Backdoor Attacks on Vision-Language-Action Models. Preprint, 2025. 5. Dexterous Grasp Generation Goal: Enable multi-fingered robotic hands to decide how to grasp an object. A grasp that is merely stable is often not the right one. Gripping a knife by the blade holds it firmly but makes it unusable, a grasp designed for a five-fingered hand may be infeasible on a four-fingered one. We therefore study generative models that produce robot grasps conditioned on what the object will be used for (task intent) and which hand is doing the grasping (hand morphology). While we investigate these factors individually, our ultimate goal is to build a single model that handles both together and works on real robot hardware. Tasks: Develop task-oriented grasp generation grounded in language or task specifications. Design morphology-aware representations that generalize across different hand embodiments. Build a unified model capable of handling both conditioning factors jointly. Conduct experiments using real robot hardware. Requirements: Proficient programming skills Experience in computer vision and deep learning (preferred) Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr) Related papers: Lee et al., DextER: Language-driven Dexterous Grasp Generation with Embodied Reasoning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026 Zhong et al., DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025 Chen et al., Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy. In Proc. Robotics: Science and Systems (RSS), 2025 6. Active Perception for Object Manipulation Goal: This project studies how a robot can actively move its wrist-mounted camera to a better viewpoint before grasping, rather than relying on a single fixed view. Most existing grasping pipelines assume a single, often fixed RGB-D observation, which frequently leaves target objects partially occluded, leading to inaccurate shape estimation and unstable grasps. From this partial observation, the system reconstructs the scene, selects the next-best camera pose expected to reveal previously occluded regions, moves the arm to acquire this view, and executes a dexterous grasp with a multi-fingered hand using the updated reconstruction. We aim to show that this active, viewpoint-aware strategy improves grasp success rate and robustness compared to single-view baselines, especially for objects with complex geometry and in cluttered scenes. Tasks: Build a simulation environment and validate grasp execution using target grasp poses. Design and implement the perception–action closed loop (reconstruction, next-best-view selection, camera move, grasp). Train and improve the next-best-view selection policy, and analyze failure cases. Extend to real robot environments. Requirements: Proficient programming skills Experience in deep learning Experience with robot simulators (preferred) Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr) Related papers: Wang, Kuancheng, et al. "Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation." arXiv preprint arXiv:2606.23625 (2026). Park, Chunghyun, Seunghyeon Lee, and Minsu Cho. "Affostruction: 3D Affordance Grounding with Generative Reconstruction." arXiv preprint arXiv:2601.09211 (2026). Lee, Junha, Eunha Park, and Minsu Cho. "Dexter: Language-driven dexterous grasp generation with embodied reasoning." arXiv preprint arXiv:2601.16046 (2026). 7. Robotic Manipulation in Cluttered Environments Goal: Improve vision-language-action (VLA) models and other robot manipulation policies to operate reliably in realistic, obstacle-rich environments. Existing benchmarks mostly cover simple pick-and-place scenarios; in real-world settings, however, obstacle placements can make the target object inaccessible without interacting with surrounding objects, requiring the robot to reason about such interactions. We aim to build a simulation-based benchmark that evaluates robot policies in such environments across different difficulty levels and use it to analyze failures and improve performance. Tasks: Construct a benchmark for robot manipulation in realistic cluttered environments. Evaluate VLA policies (e.g., OpenVLA, π0.5, GR00T N1.6) and analyze their failures. Build training data and improve VLA performance through fine-tuning. Extend to real robot environments. Requirements: Proficient programming skills Experience in deep learning Experience with robot simulators (preferred) Interested students: Reach out with a short note on your background and interests via email to Prof. Minsu Cho (mscho@postech.ac.kr). Related Papers: Physical Intelligence, π0.5: A Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054, 2025 Han et al., FetchBench: A Simulation Benchmark for Robot Fetching. In Proc. Conference on Robot Learning (CoRL), 2024 Danielczuk et al., Mechanical Search: Multi-Step Retrieval of a Target Object Occluded by Clutter. In Proc. IEEE International Conference on Robotics and Automation (ICRA), 2019
