Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands

Bench2Dex is a comprehensive simulation benchmark designed to address the lack of standardized evaluation for bimanual dexterous manipulation. While robot learning has progressed significantly, most existing benchmarks are limited to a single robot platform or a narrow set of sensing modalities. Bench2Dex fills this gap by providing a unified framework that supports 12 distinct dexterous hand embodiments, 26 long-horizon tasks—ranging from tool use to multi-stage object interaction—and approximately 1.3K human-teleoperated demonstrations.

A central challenge in dexterous manipulation is that tactile hardware varies widely in finger structure and sensor layout, making it difficult to compare policies across different hands. Bench2Dex introduces a shared visuo-tactile interface that maps local contact geometry into a consistent, image-like tactile representation. By reconstructing contact surfaces for each of the 12 supported hands and using a surface-aligned ray-casting pipeline, the benchmark allows researchers to study visuo-tactile learning in a consistent setting, regardless of the specific robot morphology.

The benchmark includes a structured evaluation suite that tests policy performance under controlled distribution shifts. It categorizes perturbations into two axes: invariance (nuisance factors like lighting or background that should not change the action) and equivariance (geometric factors like object pose that require the policy to adapt). By evaluating policies like ACT, Diffusion Policy, pi0.5, and GR00T N1.5, Bench2Dex highlights the performance gaps between matched scenes and perturbed environments, providing a clearer view of whether a policy is truly generalizing or merely memorizing training data.

Bench2Dex provides a necessary foundation for developing generalist dexterous manipulation policies. By standardizing data collection, observation modalities, and evaluation protocols across diverse embodiments, it enables the research community to move beyond single-robot benchmarks. This framework helps researchers isolate whether policy failures stem from kinematic limitations, lack of tactile awareness, or sensitivity to environmental noise, ultimately accelerating the development of more robust and versatile embodied agents.