TARMAC: A Taxonomy for Robot Manipulation in Chemistry

TL;DR

TARMAC introduces a chemistry-specific taxonomy of robot manipulation by deriving core actions from annotated teaching-lab demonstrations and validating force-dependent execution through force–torque experiments. The framework instantiates taxonomy primitives as robot-executable actions and composes them into reusable macros, bridged by the Model Context Protocol ($MCP$) to connect natural-language instructions with robotic planning. A dual validation on a chemistry manipulation dataset and a sodium chloride solution task demonstrates both descriptive coverage and practical executability, including macro synthesis via LLMs. The work lays a foundation for more scalable, reusable, and autonomous laboratory automation by offering a structured action space, shared interfaces, and future paths for richer perception and dexterous manipulation in chemical laboratories.

Abstract

Chemistry laboratory automation aims to increase throughput, reproducibility, and safety, yet many existing systems still depend on frequent human intervention. Advances in robotics have reduced this dependency, but without a structured representation of the required skills, autonomy remains limited to bespoke, task-specific solutions with little capacity to transfer beyond their initial design. Current experiment abstractions typically describe protocol-level steps without specifying the robotic actions needed to execute them. This highlights the lack of a systematic account of the manipulation skills required for robots in chemistry laboratories. To address this gap, we introduce TARMAC - a Taxonomy for Robot Manipulation in Chemistry - a domain-specific framework that defines and organizes the core manipulations needed in laboratory practice. Based on annotated teaching-lab demonstrations and supported by experimental validation, TARMAC categorizes actions according to their functional role and physical execution requirements. Beyond serving as a descriptive vocabulary, TARMAC can be instantiated as robot-executable primitives and composed into higher-level macros, enabling skill reuse and supporting scalable integration into long-horizon workflows. These contributions provide a structured foundation for more flexible and autonomous laboratory automation. More information is available at https://tarmac-paper.github.io/

Figures (8)

• Figure 1: Overview of the TARMAC framework. Laboratory manipulations are first derived from annotated teaching-lab demonstration vidoes and analyzed through feature extraction and experimental validation. These actions are then organized into the TARMAC taxonomy, which provides a structured vocabulary of primitives. The primitives can be directly executed by a robotic platform or composed into higher-level macros, which are exposed through a standardized interface (e.g., Model Context Protocol). This pipeline enables natural-language instructions descriptions to be translated into robot-executable workflows, thereby bridging human intent and robotic capability in the laboratory.
• Figure 2: Examples of annotated frames from the chemistry teaching videos. Each frame shows a representative laboratory manipulation alongside its natural-language description (top) and later corresponding taxonomic actions(bottom).
• Figure 3: Experimental setup for force–torque measurements. A Franka Research 3 robotic arm with a parallel gripper and a Bota Systems LaxONE sensor was used to manipulate laboratory glassware. The inset highlights the custom 3D-printed fixture with silicone gel padding, mounted on an aluminum frame to securely hold instruments.
• Figure 4: Experimental validation of force dependence in representative laboratory manipulations. (a) Cap insertion: Red, green, and blue lines denote forces along the $x$ ($F_x$), $y$ ($F_y$), and $z$ ($F_z$) axes. A sharp peak in $F_z$ provides a termination signal, ensuring a secure seal without glass breakage. (b) Rotational comparison: Red, green, and blue lines denote torques about the $x$ ($T_x$), $y$ ($T_y$), and $z$ ($T_z$) axes. Solid and dashed curves correspond to clamp screw tightening and tap rotation, respectively. Tap rotation shows relatively flat torque profiles governed by visual alignment, whereas screw tightening exhibits steadily increasing torque, indicating task completion. (c) Needle puncture: Red, green, and blue lines denote forces along the $x$ ($F_x$), $y$ ($F_y$), and $z$ ($F_z$) axes. Insertion into a rubber septum produces a force build-up followed by a sudden drop upon puncture—a transition undetectable by vision alone. Together, these results highlight the critical role of force–torque signals in defining reliable stopping conditions for laboratory actions.
• Figure 5: Distribution of taxonomic action labels across annotated chemistry teaching videos. Each bar represents the frequency of actions within the four major categories of TARMAC: positioning (blue), coupling (brown), tooling (green), and agitating (red). The relatively balanced counts across positioning, coupling, and tooling indicate comprehensive coverage of core manipulations, while the lower frequency of agitating actions reflects their more limited but specialized role in laboratory practice. This distribution highlights both the representativeness of the dataset and the discriminative structure of the taxonomy.