LabDex: A Hierarchical Benchmark for Dexterous Manipulation in Laboratories

Zhipeng Tang1,*, Sihang Chen1,*, Sha Zhang2,†, Peihao Yang1, Yan Liu1, Wentao Zhao1, Xinrui Lin1, Rui Huang1, Wensheng Du1, Yuting Huang1, Jiajun Deng1, Lidian Wang1, Yuan Zhang3, Yanyong Zhang1,†
1University of Science and Technology of China 2The Chinese University of Hong Kong 3iFLYTEK
*Equal contribution †Corresponding author

Front-left View

Atomic Skill, Front-left View
Compositional Skill, Front-left View
Long-horizon Workflow, Front-left View

Head View

Atomic Skill, Head View
Compositional Skill, Head View
Long-horizon Workflow, Head View

Wrist View

Atomic Skill, Wrist View
Compositional Skill, Wrist View
Long-horizon Workflow, Wrist View

Abstract

Autonomous laboratories require robots to dexterously manipulate diverse labware and execute complex, multi-stage experimental procedures. However, existing benchmarks rarely combine dexterous manipulation, real-world laboratory interactions, and long-horizon workflows within a unified framework. We introduce LabDex, a large-scale dataset and benchmark for dexterous manipulation in chemistry laboratories. LabDex unifies real-world and simulation platforms with standardized task definitions, demonstrations, and evaluation protocols, and organizes laboratory manipulation into three hierarchical levels: Atomic Skills, Compositional Skills, and Long-Horizon Laboratory Workflows. We evaluate representative robot learning methods across all three levels in both real-world and simulation environments. The results demonstrate the value of LabDex for systematically training and evaluating robotic policies, while revealing key capability bottlenecks in complex laboratory manipulation.

Overview

Overview of the LabDex real-world and simulation platforms, hierarchical task taxonomy, laboratory assets, teleoperation devices, and dataset scale.

Atomic Skills

Reusable, fine-grained primitives that capture fundamental dexterous manipulation capabilities, including precise grasping, insertion, and liquid dispensing.

Compositional Skills

Reusable laboratory operations formed by coordinating multiple atomic skills in sequence.

Long-Horizon Laboratory Workflows

Complete laboratory procedures that require robots to sequentially execute multiple compositional skills.

Data Collection

LabDex data collection pipeline for real-world and simulation environments.

Real-World Collection

LabDex collects demonstrations in both real-world and simulation environments. In the real world, a human operator teleoperates the Franka Research 3 and XHand using a VIVE Tracker and Manus Quantum data glove. During execution, demonstrations are segmented and annotated according to predefined atomic-skill boundaries, while multi-view RGB observations, robot actions, and proprioceptive states are synchronously recorded at 20 Hz.

Simulation Collection

For simulation, we extract key end-effector poses from real-world demonstrations and represent them relative to the manipulated objects. These key poses are replayed in simulation and adapted to randomized object configurations, producing simulated trajectories that preserve the same atomic-skill segmentation and hierarchical task structure as the real-world data.

Task Examples

Atomic Skills

Pick up the beaker
Pour water from the handheld beaker into the beaker on the table
Put the graduated cylinder back on the table

Compositional Skills

Place the round-bottom flask on the heater and turn it on
Pour water from the graduated cylinder into the round-bottom flask
Stir the solution in the beaker with a glass rod

Long-Horizon Laboratory Workflows

Add NaCl from the weighing funnel to the round-bottom flask

Citation

@article{tang2026labdex,
  title={LabDex: A Hierarchical Benchmark for Dexterous Manipulation in Laboratories},
  author={Tang, Zhipeng and Chen, Sihang and Zhang, Sha and Yang, Peihao and Liu, Yan and Zhao, Wentao and Liu, Xinrui and Huang, Rui and Du, Wensheng and Huang, Yuting and Deng, Jiajun and Wang, Lidian and Zhang, Yuan and Zhang, Yanyong},
  year={2026},
  eprint={2608.18618},
  archivePrefix={arXiv},
  primaryClass={cs.RO}
}