RoboTwin 2.0

'RoboTwin 2.0' is a dataset/benchmark designed to evaluate the performance of Vision-Language-Action models in generalizable robotic manipulation tasks across various unseen objects and robot embodiments.

Papers using RoboTwin 2.0 (6)

SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning2025 · 83 citesGEAR-VLA: Learning Geometry-Aware Action Representations for Generalizable Robotic Manipulation2026Agentic-VLA: Efficient Online Adaptation for Vision-Language-Action Models2026Key-Gram: Extensible World Knowledge for Embodied Manipulation2026One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies2026VLA-Thinker: Boosting Vision-Language-Action Models through Thinking-with-Image Reasoning2026