TrapVLA: Trapping Vision-Language-Action Models in Configured Failure Modes
A configured failure trapping framework for VLA models with stealthy textual triggers,
TrapEngine, TrapEval, and Trap-LIBERO / Trap-RoboTwin benchmarks.
Yi-Lin Wei 卫奕霖
Ph.D. candidate in Computer Science
I am a Ph.D. candidate at iSEE Lab, advised by Prof. Wei-Shi Zheng. I collaborated on research with the Biomimetics and Dextrous Manipulation Lab (BDML) at Stanford University.
My research focuses on robotics AI, especially dexterous grasping and manipulation. I am also interested in humanoid robotics, MLLM-driven manipulation, and dexterous hand hardware.
Recent News
- BiDexGrasp selected for a Spotlight Presentation at CoRL 2026.
- BiDexGrasp accepted at CoRL 2026.
- Humanoid Whole-Body Manipulation selected for Oral Presentation at ECCV 2026.
- Humanoid Whole-Body Manipulation accepted at ECCV 2026.
- CycleManip selected as a Highlight at CVPR 2026.
- DexGrasp-Zero accepted at RSS 2026.
Earlier news
- CycleManip and one other paper accepted at CVPR 2026.
- OmniDexGrasp accepted at ICRA 2026.
- One paper accepted in Neurocomputing.
- One paper accepted at AAAI 2026.
- TypeTele accepted at CoRL 2025.
- TacCap accepted at IROS 2025.
- AffordDexGrasp and two other papers accepted at ICCV 2025.
- I received an offer to join Stanford University as a Visiting Scholar.
- ChainHOI accepted at CVPR 2025.
- Grasp as You Say accepted at NeurIPS 2024.
- Real-to-Sim Grasp accepted at CoRL 2024.
- EconomicGrasp accepted at ECCV 2024.
- Dexterous Grasp Transformer and Single-view Scene Point Cloud Human Grasp Generation accepted at CVPR 2024.
Publications
Core contributions
Full list on Google Scholar
DynamicManip: Enabling Dynamic Manipulation from a Single Static Demonstration
A static-to-dynamic augmentation pipeline and dynamic-aware adaptive policy that enable
robots to learn dynamic manipulation from a single static demonstration.
Humanoid Whole-Body Manipulation via Active Spatial Brain and Generalizable Action Cerebellum
ECCV 2026 Oral Presentation
A generalizable humanoid loco-manipulation framework that combines Active Spatial Brain
for spatial perception and planning with Generalizable Action Cerebellum for executable
whole-body robot actions.
DexGrasp-Zero: A Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping
A morphology-aligned graph convolutional policy that learns universal dexterous grasping
from diverse embodiments, enabling zero-shot transfer to unseen robot hands with 85%
success rate.
BiDexGrasp: Coordinated Bimanual Dexterous Grasps across Object Geometries and Sizes
CoRL 2026 Spotlight Presentation
A large-scale bimanual dexterous grasping dataset and a geometry-size-adaptive grasping
generation model.
ReViP: Mitigating False Completion in Vision-Language-Action Models with Vision-Proprioception Rebalance
Mitigating false task completion in VLA models by rebalancing vision and proprioception
signals.
CycleManip: Enabling Cycle-based Manipulation via Effective History Perception and Understanding
CVPR 2026 Highlight
Achieving cyclic manipulation tasks in an end-to-end imitation manner, without relying on
auxiliary models or incurring heavy computational overhead.
OmniDexGrasp: Generalizable Dexterous Grasping via Foundation Model and Force Feedback
A generalizable dexterous framework that achieves omni-capabilities in user prompting,
dexterous embodiment, and grasping tasks.
TypeTele: Releasing Dexterity in Teleoperation by Dexterous Manipulation Types
A novel task enables robots to perform dexterous grasping based on human commands.
TacCap: A Wearable FBG-Based Tactile Sensor for Seamless Human-to-Robot Skill Transfer
A wearable FBG-based tactile sensor for transferring skills from humans to robots.
AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance
Open-Set Language-guided dexterous grasp based on generalizable-instructive
Affordance.
ChainHOI: Joint-based Kinematic Chain Modeling for Human-Object Interaction Generation
Human-Object interaction generation by joint-based kinematic chain modeling
Grasp as You Say: Language-guided Dexterous Grasp Generation
A novel task enables robots to perform dexterous grasping based on human commands.
An Economic Framework for 6-DoF Grasp Detection
A new economic grasping framework for 6-DoF grasp detection to economize the training
resource cost and meanwhile maintain effective grasp performance.
Dexterous Grasp Transformer
A novel discriminative framework for dexterous grasp generation by formulating it as a
set prediction task.
Academic Service
- Conference reviewer
- CVPR 2025-2026, NeurIPS 2024-2025, ICLR 2025-2026, ICML 2025-2026, ICCV 2025, ECCV 2026, CoRL 2025, IROS 2026.
- Journal reviewer
- RA-L.