TeleArt: Large-Scale Simulation for Zero-Shot Sim2Real Articulation Manipulation

Published:

**TeleAI, Shanghai03.2026 - 09.2026**
  • Proposed a data synthesis pipeline for articulated object manipulation and generated 1 million demonstrations for VLA pretraining. It is currently the largest synthetic dataset in the articulation manipulation domain.
  • The VLA model pretrained on the synthetic dataset outperforms the model trained from scratch and achieves performance comparable to models pretrained on large-scale real-world datasets.
  • Achieved zero-shot sim-to-real on real-world articulated object manipulation tasks across 3 robot platforms, with an average success rate of 76.7%.