Communications Chemistry · 2026

TARMAC

as a taxonomy for robot manipulation in chemistry

Kefeng Huang, Alice E. Martin, Jonathon Pipe, Tianyuan Wang, Barnabas A. Franklin,
Christopher S. Horbaczewskyj, Andy M. Tyrrell, Ian J. S. Fairlamb, Jihong Zhu

doi.org/10.1038/s42004-026-02249-3

University of York Institute for Safe Autonomy Robot-Assisted Living Laboratory (RALLA) Fairlamb Group
Two Franka robot arms beside a separating-funnel stand with a ring and a beaker

Background

A mobile robotic chemist

University of Liverpool · Burger et al., Nature 2020

What it is

Robot arm on wheels: carries samples between stations, operates instruments

Reported result

688 experiments in 8 days, a ten-variable search for photocatalysts that produce hydrogen from water

What it leaves open

One demonstrated photocatalysis workflow; touch calibration at each station

Mobile KUKA arm lowering its gripper at a laboratory station
Video: Burger et al., Nature 2020, Supplementary Video 3

Burger et al., Nature 583, 237–241 (2020), doi:10.1038/s41586-020-2442-2 · Abstract; Methods (mobile robot, touch-feedback calibration)

Background

ORGANA: a robotic chemistry assistant

Darvish et al., Matter 2025 (arXiv 2024)

What it is

Takes instructions in ordinary language, plans laboratory work, monitors experiments by camera

Reported result

A 19-step electrochemistry plan, tasks run in parallel (quinone derivatives for flow batteries)

Authors’ stated limitation

No automatic experimental setup yet, such as fetching clean vessels or inserting tubes and probes

ORGANA white robot arm above the pH-beaker station; source label Clean pH beaker
Video: authors’ project video (ac-rad.github.io/organa), 20× speed-up in source

Darvish et al., Matter 8, 101897 (2025), doi:10.1016/j.matt.2024.10.015 · arXiv:2401.06949v2, Abstract; Discussion: Limitations

Background

χDL and the Chemputer

Steiner et al., Science 2019 · Mehr et al., Science 2020

What it is

χDL writes a synthesis as process steps (add, filter, evaporate); a compiler targets compatible hardware such as the Chemputer

Reported result

12 literature compounds synthesised automatically (2020); three pharmaceuticals without human intervention (2019)

A complementary layer

Processes in abstract vessels; the physical handling can differ between apparatus

Chemputer fluidic backbone linked to reactor, filter, separator and rotary evaporator
Figure: Steiner et al., Science 2019, Fig. 3B (photograph panel)

Mehr et al., Science 370, 101–108 (2020), doi:10.1126/science.abc2986 · Steiner et al., Science 363, eaav2211 (2019), doi:10.1126/science.aav2211

Background

What these systems leave open

Execution already works
688 experiments in 8 days · a 19-step plan · 12 synthesised compounds
Physical work stays bespoke
Many manipulations (setup, adjustment, recovery, cleanup) remain manual or workflow- and platform-specific
Process is not interaction
One process operation, different manipulations across apparatus and robots
Open question
Does laboratory manipulation have a recurring physical structure?

Burger et al. (2020) · Darvish et al. (2025) · Mehr et al. (2020) · Huang et al., Commun. Chem. 2026, Introduction

TARMAC · Taxonomy

TARMAC in one picture

Laboratory instructions hide physical handling.

A taxonomy groups actions by what they physically require.

A primitive is one named action, such as Pouring or Scooping.

Each robot still needs its own controller.

Published Fig. 1: teaching-lab videos and handbooks give a set of actions, organised by three features into four categories and their primitives, composed into workflows and executed on a robot

Huang et al., Commun. Chem. 2026, Fig. 1; Introduction; Methods (Instantiating TARMAC)

TARMAC · Taxonomy

Built bottom-up from teaching videos

Corpus
91 York teaching-lab videos · 273 min
Techniques
Schlenk-line operations, reduced-pressure distillation, column chromatography, NMR sample preparation
Procedure
Watch → segment → describe, with no preset categories
Annotators
A roboticist and a chemist, jointly
Result
562 annotated atomic actions

Huang et al., Commun. Chem. 2026, Methods (Derivation of manipulation primitives); Author contributions

TARMAC · Taxonomy

Agreeing what counts as one action

One action = the smallest meaningful observable manipulation

  1. Align boundaries together.
  2. Merge synonyms.
  3. Split broad operations.
  4. Then compare how actions execute and finish.

Example: “set up the apparatus” → position a vessel, insert a component, couple two parts.

Published Fig. 2: Pull glass stopper (Transitional-C), Wipe grease (Wiping), Insert cannula stopper (Elastic-C), Rotate flask (Rotational-P)

Huang et al., Commun. Chem. 2026, Fig. 2; Methods (Derivation of manipulation primitives)

TARMAC · Taxonomy

From an observed action to a primitive

Teaching-video frame: opening a chromatography column tap
Open a column tapColumn chromatography
Teaching-video frame: pouring eluent into a chromatography column
Pour into the columnColumn chromatography
Teaching-video frame: connecting vacuum tubing to distillation apparatus
Connect vacuum tubingReduced-pressure distillation
Teaching-video frame: scooping solid for NMR preparation
Scoop a solidNMR sample preparation

Pouring · Tooling

Directing gravity-driven liquid flow by tilting or reorienting a container.

Frames: teaching-corpus examples, tarmac-paper.github.io (Dataset examples) · Definition: Huang et al., Commun. Chem. 2026, Supplementary Table 1

TARMAC · Taxonomy

Three physical distinctions

Wrench dependence
Does completion depend on force or torque?
Turn a tap to an angle; tighten a clamp until it resists.
Actuation directness
Direct contact, or through a tool or container?
Place a vessel; tilt a vessel so liquid flows.
Motion periodicity
One finite action, or sustained repeated motion?
A single press; grinding until particles are fine.

Huang et al., Commun. Chem. 2026, Results (The TARMAC taxonomy); Methods (Derivation of manipulation primitives and core features)

TARMAC · Taxonomy

Four categories of laboratory action

562 annotated actions → 4 categories → 18 primitives

Positioning

169

direct, finite, geometric endpoint

Coupling

198

direct, finite, force or torque endpoint

Tooling

137

through a tool, container or implement

Agitating

58

sustained or repeated motion

Counts are action instances in the corpus, not time spent.

Huang et al., Commun. Chem. 2026, Table 1; Fig. 3; Discussion

TARMAC · Taxonomy

18 primitives in four categories

Positioning4

Transitional-P
Setting a beaker on a benchtop
Rotational-P
Rotating a tap to adjust flow
Insertive-P
Putting a cuvette into a spectrophotometer
Sliding-P
Sliding a seal lid into position

Coupling4

Transitional-C
Pushing a cap into place
Rotational-C
Screwing the cap of a reagent bottle
Elastic-C
Fitting a rubber septum onto a flask
Penetrative-C
Piercing a septum with a needle

Tooling4

Scooping
Scooping powder with a spatula
Spotting
Applying droplets onto a TLC plate
Pouring
Pouring solvent into a beaker
Squirting
Dispensing water via a squeeze bottle

Agitating6

Shaking
Swirling a beaker to resuspend solids
Stirring
Stirring a mixture with a glass rod
Twisting
Twisting glassware to spread grease
Wiping
Wiping a vial interior with a swab
Wrapping
Wrapping parafilm around a flask neck
Grinding
Grinding a solid in a mortar and pestle

Huang et al., Commun. Chem. 2026, Supplementary Table 1 (primitives, definitions and examples from the annotated data)

TARMAC · Evaluation

Three evaluations, three different claims

EvaluationQuestionResult
Specification adequacyDo the written rules reproduce our labels?82.9–86.7% agreement on the same 562 actions
Expressive coverageDo published procedures fit the vocabulary?904 of 917 steps in 23 procedures
Robotic executabilityCan robots run and compose the primitives?8 primitives, 5 tasks, 4 configurations

Agreement, coverage and proof of concept are different measures; none is a robot success rate

Huang et al., Commun. Chem. 2026, Table 3; Introduction

TARMAC · Evaluation 1

Can the written rules reproduce our labels?

  1. GPT-5.4 receives only the written 18-primitive rulebook
  2. It labels the same 562 actions under four context conditions
  3. Its labels are compared with the original human annotation
82.9%action name only · 466 / 562
86.7%full context · 487 / 562

Macro-F1 0.74 → 0.84. A definition check on the derivation corpus, not human inter-rater agreement

0.70 0.75 0.80 0.85 0.90 Action + Procedure + Prev Full 0.829 0.851 0.861 0.867 0.74 0.77 0.83 0.84 Accuracy Macro F1

Redrawn from Huang et al., Commun. Chem. 2026, Fig. 4a; Results; Methods (Specification-adequacy probe via rule-following classification)

TARMAC · Evaluation 2

Do published procedures fit the vocabulary?

  1. 23 procedures from four published papers, not used in derivation
  2. 18 primitives fixed in advance; LLM-assisted decomposition
  3. Each paper reviewed by one of two independent chemists

904 of 917 manipulative steps mapped to a primitive (98.6%)

Positioning Coupling Tooling Agitating Paper 1 (Hydrogel mechanics) 22 5 11 10 15 3 5 4 6 4 Paper 2 (Perovskite NCs) 10 9 6 4 9 4 4 4 5 16 4 4 Paper 3 (NU-1500 MOF) 25 6 20 3 55 3 10 23 77 29 Paper 4 (Pd cross-coupling) 36 58 42 55 14 30 39 32 37 107 24 15 Transitional-P Rotational-P Insertive-P Sliding-P Transitional-C Rotational-C Elastic-C Penetrative-C Scooping Spotting Pouring Squirting Shaking Stirring Twisting Wiping Wrapping Grinding

13 unmapped: cutting or sectioning 6 · ultrasonication 3 · machine-driven tensile stretching 4

Redrawn from Huang et al., Commun. Chem. 2026, Fig. 5 (cell shading on a log scale; · marks zero); Results (Expressive coverage across published procedures); Methods (Cross-corpus protocol decomposition)

TARMAC · Evaluation 3

Can robots run the same primitives?

Single-arm Franka in a fume hood
A · Single-arm Franka: direct calls and workflow composition
Dobot arm at the bench beside its execution log
B · Dobot: nine-step solution preparation
Two Franka arms at a separating-funnel stand
C · Bimanual Franka: separating-funnel assembly
Franka with wrist camera picking up glass fragments
D · Franka + wrist depth camera: glass cleanup

8 of 18 primitives · 5 tasks · 4 configurations · 4 of 4 categories

Proof of concept, not a success-rate benchmark

Huang et al., Commun. Chem. 2026, Table 2; Fig. 6; Supplementary Videos 1–5 · clips: tarmac-paper.github.io (B and C at 4× source speed)

TARMAC · Discussion

Where TARMAC sits beside χDL and Chemputation

Protocol and process descriptions

χDL, Chemputation: which matter and energy changes are needed

↓ realised as

TARMAC: embodied actions

apparatus-dependent manipulations, named as primitives

↓ executed by

Platform-specific control

trajectories, forces, contact strategies

Adding or subtracting matter
often TARMAC actions (Pouring, Scooping, Squirting)
Adding or subtracting energy
often device-native; placing a flask on a hotplate is Transitional-P
Mapping
potentially many-to-many

Interpretation from the Discussion, not a further experiment

Huang et al., Commun. Chem. 2026, Introduction; Discussion · χDL: Mehr et al. (2020) · Chemputation: Cronin et al., PNAS (2026)

TARMAC · Discussion

Where the taxonomy stops

Neighbouring primitives
75 full-context disagreements, concentrated at the Transitional-P / Transitional-C / Insertive-P boundary
A missing primitive
Cutting or sectioning, absent from the teaching videos
A layer boundary
Ultrasonication and machine-driven stretching are device-native
Domain coverage
One teaching corpus; electrochemistry and biological workflows not targeted, photochemistry only narrowly
Robot evidence
Proof of concept; success rates and execution times were not used to evaluate the taxonomy

Huang et al., Commun. Chem. 2026, Discussion; Results; Methods

Summary

Take-home messages

  1. Laboratory manipulation has a recurring physical structure: 3 dimensions, 4 categories, 18 primitives, from 562 observed actions
  2. The fixed vocabulary expresses 904 of 917 manipulative steps in 23 external procedures; the 13 gaps mark its boundaries
  3. Selected primitives run and compose on four robot configurations

Huang et al., Commun. Chem. 2026, doi:10.1038/s42004-026-02249-3

Resources

Paper, code and data

Paper

doi.org/10.1038/s42004-026-02249-3Communications Chemistry (2026), open access, CC BY 4.0

Project website

tarmac-paper.github.ioVideos, dataset examples, overview

Code

github.com/kefhuang/tarmac-paperArchived on Zenodo: doi:10.5281/zenodo.22071890

Data

Supplementary Data 1–3 with the articleAnnotations · classification results · procedure decompositions

QR code for https://doi.org/10.1038/s42004-026-02249-3
PaperDOI link
QR code for https://tarmac-paper.github.io
Websitetarmac-paper.github.io

Huang et al., Commun. Chem. 2026, Data availability; Code availability

Acknowledgements

Thank you

Authors
Kefeng Huang, Alice E. Martin, Jonathon Pipe, Tianyuan Wang, Barnabas A. Franklin, Christopher S. Horbaczewskyj, Andy M. Tyrrell, Ian J. S. Fairlamb, Jihong Zhu
Facilities
Institute for Safe Autonomy, University of York
Funding
EPSRC: DTP studentships via the ALBERT CDT; AIchemy Frontier Fund project HATCH, EP/Y028775/1 and EP/Y028759/1; Impact Acceleration Account project SHARP, EP/X525856/1
Also supported by
Japan Society for the Promotion of Science Fellowship (PE26012)

Questions

University of York

Huang et al., Commun. Chem. 2026, Acknowledgements; Funding

Backup

Claims and evidence · feature combinations · bimanual recovery · per-primitive counts

Backup

Claims supported by each evaluation

Evaluation exerciseSpecific claim supported
Rule-following classificationConcordance of 82.9–86.7% and macro-F1 of 0.74–0.84 across 562 actions show that the written specification recovers intended assignments within the derivation corpus and that workflow context contributes to primitive identification.
Cross-corpus protocol decompositionMapping 904 of 917 manipulative steps across 23 procedures, while retaining 13 unmapped steps as boundary cases, supports high expressive coverage within the examined external procedures while exposing operation classes outside the present scope.
Robotic instantiationFive tasks across four configurations, covering 8 of 18 primitives and all four high-level categories, provide proof-of-concept evidence that representative primitives can be instantiated, composed, transferred across embodiments, and applied across the experimental lifecycle.

Redrawn from Huang et al., Commun. Chem. 2026, Table 3

Backup

Feature combinations behind the four categories

Wrench DependentIndirect ControlCyclic MotionCount ↓Taxonomic Category
✓✗✗198Coupling
✗✗✗169Positioning
✗✓✗73Tooling
✓✓✗64
✗✗✓44Agitating
✗✓✓13
✓✗✓0
✓✓✓1

Highlighted cells mark the dominant feature of each category. Counts are annotated actions (total 562).

Redrawn from Huang et al., Commun. Chem. 2026, Table 1

Backup

Bimanual recovery by recomposing primitives

Published Fig. 6 row C: Screwing (Rotational-C), Loosen for recovery, Re-rotate ring flat, Pour DCM (Tool.)

Misaligned iron ring: Rotational-C (loosen) → Rotational-P (re-orient) → Rotational-C (re-secure), then assembly continues.

Huang et al., Commun. Chem. 2026, Fig. 6 (row C); Results (Robotic instantiation); Supplementary Video 4

Backup

Annotated actions per primitive

Positioning169 transitional (p) 33 rotational (p) 53 insertive (p) 71 sliding (p) 12 Coupling198 transitional (c) 52 rotational (c) 89 penetrative (c) 17 elastic (c) 40 Tooling137 squirting 64 pouring 57 scooping 13 spotting 3 Agitating58 wiping 24 shaking 12 stirring 1 twisting 12 wrapping 8 grinding 1

Redrawn from Huang et al., Commun. Chem. 2026, Fig. 3 (562 annotated actions in the teaching-video corpus)