Source-linked AI summary
ORGANA: A Robotic Assistant for Automated Chemistry Experimentation and Characterization
Kourosh Darvish, Marta Skreta, Yuchi Zhao, Naruki Yoshikawa, Sagnik Som, Miroslav Bogdanovic, Yang Cao, Han Hao, Haoping Xu, Alán Aspuru-Guzik, Animesh Garg, Florian Shkurti
TL;DR
Chemistry material discovery remains time- and labor-intensive, with gaps between specialized laboratory workstations limiting automation. ORGANA combines language-based interaction, visual perception, decision-making, planning, and scheduling to assist chemists across experiments, reproducing literature results and improving parallel execution time by 21.1%.
Problem
Material discovery remains time- and labor-intensive, while chemists must bridge automation gaps between specialized laboratory workstations.
Method
ORGANA uses language-based interaction, visual perception, decision-making, planning, and scheduling to assist chemists across diverse chemistry experiments.
Results
Organa reproduced literature results across experiments and improved overall execution time by 21.1% through parallel planning and scheduling.
Takeaways & Limitations
ORGANA provides a flexible assistive approach for automating and coordinating fundamental chemistry experiments with reduced manual involvement.
Takeaways & Limitations
ORGANA relies primarily on independent sensor modalities and lacks online replanning, limiting perception and adaptation under execution uncertainty.
Abstract
from arXiv · showhide
Chemistry experiments can be resource- and labor-intensive, often requiring manual tasks like polishing electrodes in electrochemistry. Traditional lab automation infrastructure faces challenges adapting to new experiments. To address this, we introduce ORGANA, an assistive robotic system that automates diverse chemistry experiments using decision-making and perception tools. It makes decisions with chemists in the loop to control robots and lab devices. ORGANA interacts with chemists using Large Language Models (LLMs) to derive experiment goals, handle disambiguation, and provide experiment logs. ORGANA plans and executes complex tasks with visual feedback, while supporting scheduling and parallel task execution. We demonstrate ORGANA's capabilities in solubility, pH measurement, recrystallization, and electrochemistry experiments. In electrochemistry, it executes a 19-step plan in parallel to characterize quinone derivatives for flow batteries. Our user study shows ORGANA reduces frustration and physical demand by over 50%, with users saving an average of 80.3% of their time when using it.
INTRODUCTION
Organa is an assistive robotic system for flexible chemistry-lab automation that translates chemists’ instructions into executable experiments using LLMs, perception, decision-making, and human feedback. It combines task-and-motion planning with scheduling to execute chemistry tasks in parallel while supporting multiple experiments, troubleshooting, and automated reporting.
- 3D visual perception lets Organa manipulate laboratory objects, monitor experiments, and decide when robots should proceed to subsequent steps.The system combines autonomous decision-making with high-level human involvement when needed.
- Organa translates chemists’ audio or text commands into experiment descriptions and robot goals using an LLM-based reasoning architecture.It also grounds perceived objects through user interaction and supports timely human intervention.
- Organa demonstrates flexible automation across fundamental chemistry experiments and electrochemical characterization of quinone derivatives.The introduction presents solubility, recrystallization, and pH testing as example applications, alongside electrochemical-property characterization.
- Organa simultaneously solves task-and-motion planning and scheduling, enabling multiple robots and devices to execute chemistry tasks in parallel.This joint formulation is intended to improve experimentation efficiency and laboratory-resource utilization.
- The system supports multiple experiments, troubleshooting unexpected behavior and ambiguities, and feedback through reports summarizing plans, results, analyses, errors, and user-provided information.Unlike CLAIRify’s single-experiment reasoning, Organa can plan experiment series over ranges of parameters.
RESULTS
Organa reproduced diverse multistep chemistry experiments and demonstrated electrochemical characterization with comparable results to chemists. Parallel execution reduced electrochemistry time by 21.1%, while user studies found substantially lower interaction burden and physical demand.
- RESULTS: Organa was evaluated across solubility screening, recrystallization, pH testing, and electrochemistry experiments to assess reliability and modularity.The first three experiments were reproduced from prior work, while the results focus on advanced electrochemistry and user interaction.
- Electrochemistry Experiment: 21.1%: Parallel electrochemistry execution reduced average experiment time from 21.67 to 17.10 minutes versus sequential execution.Sequential planning took 61.52±0.1 s, while temporal task and motion planning took 186.3±46.0 s over 12 trials.
- User study – quantitative results: Organa produced electrochemical estimates comparable to chemists, with pKa1 values of 8.03 versus 8.02 and slopes of -61.3 versus -62.7 mV/pH unit.The comparison used combined data from all human experiments and all Organa experiments.
- User study – quantitative results: Organa required 7.35 minutes for written instructions and 4.27 minutes for spoken instructions during startup, compared with over 30 minutes for manual experimentation.Troubleshooting required an average of 1.30 minutes of chemist feedback.
- User study – qualitative analysis: Organa halved participant frustration and reduced physical demand fourfold compared with manual experimentation.NASA-TLX responses showed reduced demand and effort across all measured categories.
DISCUSSION
ORGANA reliably reproduces chemistry results, supports modular and parallel experimentation, and reduces users’ physical and temporal workload while retaining human oversight. Its current limitations include restricted perception, difficult planner-domain modification, no online replanning, and absent automatic setup preparation.
- Reliability and reproducibility of chemistry results: Organa reproduced literature and chemist results, estimating solubility at 10.2 ± 2.2% and electrochemical values of −61.4 ± 0.5 mV /pH unit and 8.03 ± 0.17 pKa1.The electrochemistry values were obtained across three runs and were comparable to chemists’ results.
- Modularity: Organa’s modular architecture integrates NLP, perception, TAMP, robot execution, data analysis, and laboratory hardware across varied chemistry experiment setups.It jointly solves task-and-motion planning and scheduling to parallelize available pumps, arms, stirrers, polishers, potentiostats, and pH sensors.
- Interactions Between Chemists and Organa: 88.4% of users’ time was saved using audio interaction with Organa versus manual experimentation for three buffer solutions.Figure 7 quantifies the reduction in human temporal workload, despite subjective studies not finding a significant efficiency increase.
- Evaluation of autonomy and robustness: Organa completed four long-horizon experiments, including solubility, recrystallization, pH measurement, and electrochemistry, with plans ranging from 6 to 114 steps.Execution times were 25.63, 44.80, 3.85, and 130.00 minutes, respectively.
- Evaluation of autonomy and robustness: Organa’s human-in-the-loop design improves robustness by detecting issues, informing users, and enabling timely corrections during experiments.The system also uses constrained motion planning, consistency checks, and feedback integration to address laboratory safety.
- Interactions Between Chemists and Organa: Users reported reduced physical load and frustration, usefulness across CLI, GUI, and natural-language interfaces, and continued need for humans in the autonomy loop.Half of participants were uncertain about trusting a robot to complete experiments autonomously, while comprehensive reports and uncertainty notifications could increase trust.
- Limitations: Organa remains limited by independent sensor modalities, complex PDDL-domain definition, absent online replanning, and inability to automatically prepare experimental setups.Future setup automation could include retrieving vessels, inserting tubes or probes, and then running experiments.
EXPERIMENTAL PROCEDURES · Resource availability · Data and code availability:
ORGANA combines LLM-based reasoning, human-in-the-loop clarification, scheduled parallel planning, perception, chemistry-specific robot skills, and automated analysis to execute and document experiments. Its methods support adaptive multi-experiment planning, workspace understanding, material monitoring, and electrochemical parameter estimation and reporting.
- EXPERIMENTAL PROCEDURES: ORGANA.Reasoner converts chemists’ high-level instructions into structured experiment goals, incorporates past outcomes and user feedback, and adapts plans across interdependent experiments.CLAIRify converts natural-language descriptions into syntactically valid XDL codes for planning, while unexpected results trigger user verification and revised reasoning.
- Data and code availability:: The system architecture links user intention grounding, LLM-based reasoning, parallel task and motion planning, and robot execution with action and perception skills.Figure 9 presents the overall workflow connecting these components for chemistry experimentation.
- Resource availability: ORGANA uses human-in-the-loop scene grounding to identify vessel functions and resolve ambiguities before planning actions in workspaces containing similar objects.Users label vessels by intended use, such as reaction or waste, and can interact through text or speech.
- EXPERIMENTAL PROCEDURES: ORGANA’s planner schedules parallel robot and equipment actions using durative PDDL2.1 actions, time-variant costs, weighted A* search, and precondition checks that prevent race conditions.PDDLStream actions are transformed into start and end actions, enabling concurrent plans and feasible scheduling under experiment constraints.
- EXPERIMENTAL PROCEDURES: ORGANA’s dual-level perception monitors chemical progress and workspace objects using sensors, turbidity estimation, and transparent-object detection and pose estimation.Average solution brightness provides a turbidity proxy, while Grounding DINO, depth data, point-cloud processing, and geometric estimation support object perception.
- EXPERIMENTAL PROCEDURES: Chemistry-specific skills cover manipulation, constrained transport, liquid and powder pouring, and electrode polishing for solubility, recrystallization, and electrochemistry experiments.The system uses pre/post manipulation poses, orientation constraints, PD-controlled weight feedback, and a mechanically compliant polishing station.
- Data and code availability:: ORGANA automatically generates a PDF experiment report containing experiment details, statistical analyses, failure logs with resolutions, and summary plots.The report is produced at the conclusion of the experiment and includes automatically generated analyses of electrochemical measurements.
- EXPERIMENTAL PROCEDURES: For electrochemistry, ORGANA models pH–redox-potential behavior with three linear regions and estimates parameters using maximum likelihood and sequential posterior updates.Posterior distributions and model lines are updated after each sampled data point to keep chemists informed about experiment progress.
Supplemental information index
The supplemental materials provide experimental details, related figures, a table, and a video demonstrating ORGANA’s parallel electrochemistry execution with human interaction.
- The supplemental PDF includes notes on experimental details.
- It also contains corresponding figures and a table.
- Video S1 demonstrates ORGANA interacting with a human while executing an electrochemistry experiment in parallel.
Supplemental information
The supplemental information details ORGANA’s perception pipeline, reasoning and planning implementations, Bayesian parameter estimation, and an example quinone experiment. Perception achieved 3.5 cm mean position-estimation error after point-cloud outlier removal, while detection performance varied by object and IoU threshold.
- Perception Analysis: Glass-object detection maintained consistently high AP across IoU thresholds, whereas plate detection declined sharply as IoU increased.The decline occurred because Grounding DINO often recognized the brown pad as a plate.
- Perception Analysis: 3.5 cm MAE was achieved for object position estimation after radius outlier removal improved performance from 4.5 cm to 3.5 cm.The small beaker had the lowest MAE at 2.4 cm, while the large flask had the highest at 5.1 cm.
- Perception Analysis: The perception pipeline combines Grounding DINO and SAM for prompted detection and segmentation with depth-based point-cloud filtering for object-position estimation.Grounding DINO uses “glass object” and “plate” prompts, while the evaluated checkpoints use Swin-T and ViT-B backbones.
- Organa.Reasoner Implementation and Prompts: ORGANA’s planning implementation uses PDDLStream to combine predicates, actions, initial and goal states, conditional samplers, and integrated task-and-motion planning.PDDL standardizes reusable and comparable planning representations, while PDDLStream supports concurrent task execution through scheduling.
- Experiment Summary: The supplemental experiment summary describes quinone potential measurements from pH 7 to pH 9 using buffer, sodium chloride, water, and quinone solution before cyclic-voltammetry scans.After each scan, the beaker contents were transferred to a waste beaker.