BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks
Yong Wang, Kecheng Zheng, Xing Zhu, Chao Xu, Yujun Shen, Shengze Cai
Zhejiang University · Ant Group
Published October 18, 2026 · on the feed October 21, 2026 · ▲ 1 · arXiv 2610.25459
The abstract
Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models insufficiently characterized. In this paper, we present BiDex-100, a simulation benchmark for bimanual dexterous manipulation that combines broad task coverage, multi-dimensional evaluation, and high-DoF hand embodiments. Concretely, the benchmark comprises 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes, allowing systematic assessment of robustness. In addition, the benchmark supports two embodiments and provides 8,080 success-validated teleoperated demonstrations, totaling 6.25 million frames. We evaluate seven recent Vision-Language-Action models under a unified protocol, establishing baselines for capability-specific performance, robustness, and data efficiency. The website is available at https://BiDex-100.github.io/BiDex-100/.
The Brief
2:08A short podcast summary of the paper. Written and voiced by AI.
Transcript
BiDex-100. A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks. By Yong Wang and five others, from Zhejiang University and Ant Group. Published on the eighteenth of October. Its code repository is worked out: about sixty-five source files, with a readme and a licence.
Bimanual manipulation with dexterous hands, two hands with many moving parts working together, has drawn growing attention in research on robotic action models. Yet few studies have built benchmarks tailored to dexterous manipulation in particular. As a result, how well existing models handle dexterous manipulation has been insufficiently characterized.
BiDex-100 is a simulation benchmark built for that setting. It measures performance along five capability axes: generalization, instruction following, working from limited demonstrations, long-horizon tasks, and memory. On top of those five, it measures robustness, through an out-of-domain variant of every task.
The benchmark comprises one hundred and six tasks organized along those axes. Each task is paired with its out-of-domain variant, which introduces controlled visual and spatial changes, so robustness can be assessed systematically rather than anecdotally. Two hand embodiments with high degrees of freedom are supported.
Alongside the tasks, the benchmark provides eight thousand and eighty teleoperated demonstrations, each one validated as successful. Together they total six point two five million frames.
Seven recent Vision-Language-Action models are evaluated under one unified protocol, so their numbers can be set side by side. The outcome is a set of baselines in three parts: performance on each specific capability, robustness when a task shifts to its out-of-domain variant, and data efficiency. The three are set out as separate baselines, not folded into a single overall figure.
What BiDex-100 contributes is breadth and a common yardstick. Broad task coverage, evaluation along several dimensions, and high-DoF hand embodiments are combined in one place, so a model's dexterity can be read as a profile across capabilities rather than as a single score.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free