CoChat

Large language models for chemistry robotics

Open access

The DOI matches its Crossref record; 108 of the 109 references it cites were found in a registry or an index, and 1 are not checked yet.

Abstract

This paper proposes an approach to automate chemistry experiments using robots by translating natural language instructions into robot-executable plans, using large language models together with task and motion planning. Adding natural language interfaces to autonomous chemistry experiment systems lowers the barrier to using complicated robotics systems and increases utility for non-expert users, but translating natural language experiment descriptions from users into low-level robotics languages is nontrivial. Furthermore, while recent advances have used large language models to generate task plans, reliably executing those plans in the real world by an embodied agent remains challenging. To enable autonomous chemistry experiments and alleviate the workload of chemists, robots must interpret natural language commands, perceive the workspace, autonomously plan multi-step actions and motions, consider safety precautions, and interact with various laboratory equipment. Our approach, CLAIRify , combines automatic iterative prompting with program verification to ensure syntactically valid programs in a data-scarce domain-specific language that incorporates environmental constraints. The generated plan is executed through solving a constrained task and motion planning problem using PDDLStream solvers to prevent spillages of liquids as well as collisions in chemistry labs. We demonstrate the effectiveness of our approach in planning chemistry experiments, with plans successfully executed on a real robot using a repertoire of robot skills and lab tools. Specifically, we showcase the utility of our framework in pouring skills for various materials and two fundamental chemical experiments for materials synthesis: solubility and recrystallization. Further details about CLAIRify can be found at https://ac-rad.github.io/clairify/ .

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1007/s10514-023-10136-2

Cites

108 of 109 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Task plans generated by LLMs may contain syntactical errors in domain-specific languages. By using verifier-assisted iterative prompting, CLAIRify can generate a valid program. Once the program has been verified, it is passed on to a task and motion planner (TAMP) for execution by a robot
Fig. 1 Task plans generated by LLMs may contain syntactical errors in domain-specific languages. By using verifier-assisted iterative prompting, CLAIRify can generate a valid program. Once the program has been verified, it is passed on to a task and motion planner (TAMP) for execution by a robotCC-BY · From the PDF
Fig. 2 Our framework. a CLAIRify : LLM-based natural language processing module. The LLM takes the input (1), structured language definition, and (optionally) resource constraints and generates unverified structured language (2). The output is examined by the verifier and is passed to LLM with feedback (3). The LLM-generated outputs pass through the verifier (4). The correct output (5) is passed to the task and motion planning module (6) to generate robot trajectories. b Robot planning module, which is composed of Perception , Task and Motion Planning , and Skills blocks. Our framework enables the robot to leverage available chemistry lab devices (including sensors and actuators) by
Fig. 2 Our framework. a CLAIRify : LLM-based natural language processing module. The LLM takes the input (1), structured language definition, and (optionally) resource constraints and generates unverified structured language (2). The output is examined by the verifier and is passed to LLM with feedback (3). The LLM-generated outputs pass through the verifier (4). The correct output (5) is passed to the task and motion planning module (6) to generate robot trajectories. b Robot planning module, which is composed of Perception , Task and Motion Planning , and Skills blocks. Our framework enables the robot to leverage available chemistry lab devices (including sensors and actuators) byCC-BY · From the PDF
Fig. 4 Web interface for CLAIRify . Users input natural language descriptions of the experiment in the left column. XDL is generated in the right column when the user pushes the Translate button
Fig. 4 Web interface for CLAIRify . Users input natural language descriptions of the experiment in the left column. XDL is generated in the right column when the user pushes the Translate buttonCC-BY · From the PDF
Image 4
CC-BY · From the PDF
Fig. 5 Pouring skill controller: given the XDL and TAMP reference values and sensor feedback, the pouring controller computes the endeffector velocity for the robot by blending a shaping function s ( t ) and a PD control output v PD ( t )
Fig. 5 Pouring skill controller: given the XDL and TAMP reference values and sensor feedback, the pouring controller computes the endeffector velocity for the robot by blending a shaping function s ( t ) and a PD control output v PD ( t )CC-BY · From the PDF
Fig. 6 An example of pouring control. The velocity of the end-effector is controlled based on the feedback error and shaping function
Fig. 6 An example of pouring control. The velocity of the end-effector is controlled based on the feedback error and shaping functionCC-BY · From the PDF
Fig. 7 An example of automated turbidity measurement. The camera detects the Petri dish using Hough Circle Transform. The average brightness of the detected area (red square) is used as a proxy of turbidity (Color figure online)
Fig. 7 An example of automated turbidity measurement. The camera detects the Petri dish using Hough Circle Transform. The average brightness of the detected area (red square) is used as a proxy of turbidity (Color figure online)CC-BY · From the PDF
Fig. 8 Violin plots showing distributions of different error categories in XDL plans generated for experiments for the Chem-RnD (left) and Chem-EDU (right) datasets. The x-axis shows the error categories and the y-axis shows the number of errors for that category (lower is better). For the Chem-RnD dataset, we show the error distributions for both CLAIRify and SynthReader. Each violin is split in two, with the left
Fig. 8 Violin plots showing distributions of different error categories in XDL plans generated for experiments for the Chem-RnD (left) and Chem-EDU (right) datasets. The x-axis shows the error categories and the y-axis shows the number of errors for that category (lower is better). For the Chem-RnD dataset, we show the error distributions for both CLAIRify and SynthReader. Each violin is split in two, with the leftCC-BY · From the PDF
Fig. 9 Feedback loop between the Generator and Verifier. The input text is converted to structured-like language via the generator and is then passed through the verifier. The verifier returns a list of errors (marked with a yellow 1). The feedback is passed back to the generator along with the erroneous task plan, generating a new task plan. Now
Fig. 9 Feedback loop between the Generator and Verifier. The input text is converted to structured-like language via the generator and is then passed through the verifier. The verifier returns a list of errors (marked with a yellow 1). The feedback is passed back to the generator along with the erroneous task plan, generating a new task plan. NowCC-BY · From the PDF
Image 10
CC-BY · From the PDF
Image 11
CC-BY · From the PDF
Fig. 12 Workflow of solubility experiment. a We translate a natural language input to XDL using CLAIRify . b We then execute the plan on a robot using TAMP. First, a fixed amount of the solute is added to the dish on the weighing scale and stirrer. The robot pours 10 g of water
Fig. 12 Workflow of solubility experiment. a We translate a natural language input to XDL using CLAIRify . b We then execute the plan on a robot using TAMP. First, a fixed amount of the solute is added to the dish on the weighing scale and stirrer. The robot pours 10 g of waterCC-BY · From the PDF
Fig. 13 Turbidity change during experiment. Water is poured into the dish during pouring (grey) and turbidity is measured during observation (blue). The end of the experiment is determined by turbidity comparison. In this example, all solutes are dissolved at the third pouring because the turbidity change after the fourth pouring is below the threshold. The average weight of the second and third pouring is used to calculate the solubility (Color figure online)
Fig. 13 Turbidity change during experiment. Water is poured into the dish during pouring (grey) and turbidity is measured during observation (blue). The end of the experiment is determined by turbidity comparison. In this example, all solutes are dissolved at the third pouring because the turbidity change after the fourth pouring is below the threshold. The average weight of the second and third pouring is used to calculate the solubility (Color figure online)CC-BY · From the PDF
Figure 14 shows the result of the experiment.
Figure 14 shows the result of the experiment.CC-BY · From the PDF
Image 15
CC-BY · From the PDF
Image 16
CC-BY · From the PDF
Image 17
CC-BY · From the PDF
Image 18
CC-BY · From the PDF
Image 19
CC-BY · From the PDF
Image 20
CC-BY · From the PDF
Image 21
CC-BY · From the PDF
Image 22
CC-BY · From the PDF
Image 23
CC-BY · From the PDF
Image 24
CC-BY · From the PDF
Image 25
CC-BY · From the PDF
Image 26
CC-BY · From the PDF
Image 27
CC-BY · From the PDF
Image 28
CC-BY · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 109 references

Research works printed but not linked: 1

  1. Yoshikawa, N., Li, A. Z., Darvish, K., Zhao, Y., Xu, H., Kuramshin, A., Aspuru-Guzik, A., Garg, A., & Shkurti, F. (2023). Chemistry lab automation via constrained task and motion planning. arXiv preprint. https://doi.org/10.48550/arXiv.2212.09672 · DOI

Cited by

See all 130

Metadata

Authors
Yoshikawa, N., Skreta, M., Darvish, K., Arellano-Rubach, S., Ji, Z., Bjørn Kristensen, L., Li, A. Z., Zhao, Y., Xu, H., Kuramshin, A., Aspuru-Guzik, A., Shkurti, F., & Garg, A.
Affiliations
  • University of Toronto (CA)
  • University of Waterloo (CA)
Corresponding author
Darvish, Kourosh
Venue
Autonomous Robots
Published
2023-10-25
Volume
47
Issue
8
Pages
1057-1086
ISSN
  • 0929-5593 (ISSN-L)
  • 1573-7527
OpenAlex
W4387937021
Cited by
112
References
94
Citations by year
2026: 39 · 2025: 43 · 2024: 30
FWCI
9.91
Type
journal article
Language
en
Open access
Licence
Locations
Autonomous Robots (publishedVersion, cc-by)
Funders
Carlsberg Foundation (CF21-0669)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
  • Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
  • Physical Sciences › Engineering › Control and Systems Engineering › Robotics and Automated Systems
Keywords
large language models, natural language interfaces, task and motion planning, iterative prompting, program verification, embodied agents, laboratory automation, recrystallization, collision avoidance

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Large language models for chemistry robotics. https://app.cochat.ai/papers/10.1007/s10514-023-10136-2
The paper
Yoshikawa, N., Skreta, M., Darvish, K., Arellano-Rubach, S., Ji, Z., Bjørn Kristensen, L., Li, A. Z., Zhao, Y., Xu, H., Kuramshin, A., Aspuru-Guzik, A., Shkurti, F., & Garg, A. (2023). Large language models for chemistry robotics. Autonomous Robots. https://doi.org/10.1007/s10514-023-10136-2

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.