GenDexHand: Generative Simulation for Dexterous Hands

Overview

GenDexHand is a generative simulation pipeline for dexterous hands. It turns task ideas into scalable, high-quality trajectory data by combining language and vision models with motion planning, degree-of-freedom constraints, and task decomposition.

From task ideas to learning

• Generate tasks: synthesize dexterous manipulation scenarios.
• Refine the scene: use VLM feedback to adjust object placements and scales.
• Break down the task: decompose complex tasks into subtasks for sequential reinforcement learning.

My contribution

I developed the generative simulation pipeline, connecting task synthesis, visual refinement, and sequential learning for dexterous hands.