ARMS: an Automated Robotized human-Mimicking Solid sampling system
- 1Swiss Cat+ West Hub
- 2École Polytechnique Fédérale de Lausanne (EPFL
- 3Idiap Research Institute
Abstract
Chemists sample solid materials such as powders, highly viscous fluids, and semi-solids with a spatula in labs every day. Yet, this action remains the key bottleneck in the development of fully automated chemistry laboratories. While a handful of specialist tools have been developed to autonomously manipulate solids, these approaches target a limited range of material properties. In this work, we approach the problem from the other perspective by studying the capabilities of chemists to gain insights into the design of a generalist solid sampling approach with the aim of spanning the entire range of possible material properties. We observed and demonstrated that the key capabilities are characterizing, identifying the material descriptors that map material properties to actions; scooping, adding material to the spatula; depositing, removing material from the spatula; weighing, measuring the mass transferred; and auxiliary actions such as cleaning the spatula. We demonstrate how each of these actions can be combined in a bimanual robot platform and sample seven of nine materials chosen to span the range of possible materials. In order to achieve this performance, we leverage low-data learning from demonstration techniques such that demonstration collection takes less than an hour and training lasts less than a minute, exploit geometric structure, and transfer to novel materials within a descriptor class. Notably, we were able to move not only simpler granular free-flowing materials such as NaCl but also sticky materials such as honey, reaching standard deviations of 1.5–3.3 mg and 2.8–10.1 mg with scoop depth predicting delivered mass at 0.70 < R2 < 0.93 and 0.28 < R2 < 0.66, respectively.
Related
- An Autonomous Robotized Cell for Polymer Foaming
- Poly(lactic acid) nanoplastics through laser ablation: establishing a reference model for mimicking biobased nanoplastics in aquatic environments
- Triple Helical-Peptides: Collagen-Mimicking Tools for Tissue Engineering
- Screw dislocation–mediated galaxy-mimicking superspiral assembly with amplified bio-piezoelectricity
- United Complexes or [2]Catenanes: Curvature-Controlled Cooperative π-Coordination Mimicking θ-Curve Motifs
BibTeX
@misc{tan2026arms,
title = {{ARMS: an Automated Robotized human-Mimicking Solid sampling system}},
author = {Jasper Tan and James Hermus and Emmanuel Senft and Sylvain Calinon and Pascal Miéville},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009421/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009421/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - ARMS: an Automated Robotized human-Mimicking Solid sampling system
AU - Jasper Tan
AU - James Hermus
AU - Emmanuel Senft
AU - Sylvain Calinon
AU - Pascal Miéville
PY - 2026
DA - 2026/09/24
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009421/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009421/v1
AB - Chemists sample solid materials such as powders, highly viscous fluids, and semi-solids with a spatula in labs every day. Yet, this action remains the key bottleneck in the development of fully automated chemistry laboratories. While a handful of specialist tools have been developed to autonomously manipulate solids, these approaches target a limited range of material properties. In this work, we approach the problem from the other perspective by studying the capabilities of chemists to gain insights into the design of a generalist solid sampling approach with the aim of spanning the entire range of possible material properties. We observed and demonstrated that the key capabilities are characterizing, identifying the material descriptors that map material properties to actions; scooping, adding material to the spatula; depositing, removing material from the spatula; weighing, measuring the mass transferred; and auxiliary actions such as cleaning the spatula. We demonstrate how each of these actions can be combined in a bimanual robot platform and sample seven of nine materials chosen to span the range of possible materials. In order to achieve this performance, we leverage low-data learning from demonstration techniques such that demonstration collection takes less than an hour and training lasts less than a minute, exploit geometric structure, and transfer to novel materials within a descriptor class. Notably, we were able to move not only simpler granular free-flowing materials such as NaCl but also sticky materials such as honey, reaching standard deviations of 1.5–3.3 mg and 2.8–10.1 mg with scoop depth predicting delivered mass at 0.70 < R2 < 0.93 and 0.28 < R2 < 0.66, respectively.
ER -