Bench2Dex
Bench2Dex research collaborators
A released Isaac Lab benchmark for simulated bimanual visuo-tactile manipulation. The authors report 12 hand embodiments, 26 tasks and about 1,300 teleoperated demonstrations.
What you can explore
Explore the published code, teleoperation files and executable task evaluation. Pin the release and follow the documented preprocessing before training.
Before you use it
Rendered tactile observations do not reproduce a particular physical sensor. Tasks are not crossed with every hand, so the results do not isolate hand-design effects. The dataset viewer failed during review; use the original files and validate a sample locally. Dexhands has not reproduced the results.
Original contributors
Zhenjie Yang, Yideng Zhang, Dongjie Zhang, Chenyu Jiang and collaborators
License / access: Code: MIT. Dataset, model and robot-asset permissions must be checked separately; the reviewed dataset repository had no dataset card.
Research overview and limitations
Editorial reference · checked 2026-09-22