RoboTwin

Common

A bimanual-manipulation simulation data generator and evaluation benchmark from the University of Hong Kong, Shanghai AI Lab, and others.

RoboTwin is a simulation data-generation and evaluation platform for bimanual manipulation, released by more than a dozen institutions including the University of Hong Kong's MMLab, Shanghai AI Lab, Tsinghua, and Shanghai Jiao Tong University, built on the SAPIEN simulator; version 1.0 received a CVPR 2025 Highlight. Version 2.0, released in June 2025, has a multimodal large model automatically write expert manipulation code, which is verified in simulation and then used to generate trajectories, covering 50 bimanual tasks, 5 robot embodiments, and 731 objects, with over 100,000 trajectories pre-collected. It applies domain randomization (randomly varying environmental conditions during training) across five dimensions — clutter, background texture, lighting, table height, and language instructions — to improve real-robot robustness, and it also serves as a common bimanual evaluation benchmark for VLA models.

ExampleThe RoboTwin 2.0 paper reports that a VLA model trained on its heavily randomized data achieved a 367% relative improvement in previously unseen real-world scenes, and the CVPR 2025 MEIS workshop ran a challenge built on it.

Also called: RoboTwin 2.0
Related: Bimanual Manipulation · Domain Randomization · Synthetic Data · SAPIEN (SimulAted Part-based Interactive ENvironment) · Benchmark · Cross-Embodiment
Sources: RoboTwin 官网 (Chinese)
RoboTwin 2.0 (arXiv 2506.18088)
As of: 2025-06