Projects
We are working with renowned labs and start-ups. Our research was published in top-tier journals.

Scientific Publications

Selected papers we co-authored across chemistry, lab automation, AI and robotics.

1. A nanomaterials discovery robot for the Darwinian evolution of shape programmable gold nanoparticles Nature Communications2020D Salley, G Keenan, J Grizou, A Sharma, S Martín, L Cronin

   A robot evolved gold nanoparticles across generations, passing the best results of one round back in as the physical seeds of the next.

   The platform prepared and analysed its own reactions using in-line UV-Vis spectroscopy, while a genetic algorithm chose the next set of conditions. It did not only pass digital parameters between generations: it reused the nanoparticles themselves as seeds, so synthetic history shaped later discovery. The system learned conditions for spheres, then rods, then used the optimised rods to reach more complex octahedral particles.

   Read the paper
2. Networking chemical robots for reaction multitasking Nature Communications2018D Caramelli, D Salley, A Henson, G A Camarasa, S Sharabi, G Keenan, L Cronin

   Low-cost chemistry robots were networked so they could share experimental choices and results in real time.

   Instead of treating a robot as an isolated machine, the work connected several so they could explore reaction spaces together. The networked set synchronised oscillating reactions, encoded information chemically, and assessed how reproducible crystallisations were. The lesson carries directly into modern automated labs: hardware, software, data and coordination need designing together.

   Read the paper
3. Adaptive artificial evolution of droplet protocells in a 3D-printed fluidic chemorobotic platform with configurable environments Nature Communications2017J M Parrilla-Gutierrez, S Tsuda, J Grizou, J Taylor, A Henson, L Cronin

   A 3D-printed robotic platform evolved droplet protocells while changing the environment they lived in.

   The system generated and selected droplets in real time, with the surrounding environment treated as an experimental variable rather than a fixed background. Making the platform by rapid prototyping is what allowed the environment to be reconfigured between runs. The work showed the environment acting as an active selector on which droplets adapt.

   Read the paper
4. An artificial intelligence enabled chemical synthesis robot for exploration and optimization of nanomaterials Science Advances2022Y Jiang, D Salley, A Sharma, G Keenan, M Mullin, L Cronin

   A synthesis robot combined automated experiments with machine learning to explore and optimise nanomaterials.

   The platform ran the full cycle end to end: preparing samples, measuring them, and using the results to choose what to try next. It is the larger sibling of the nanomaterials evolution work, built to search a wider space of conditions.

   Read the paper
5. A curious formulation robot enables the discovery of a novel protocell behavior Science Advances2020J Grizou, L J Points, A Sharma, L Cronin

   Instead of optimising toward a target, a curiosity algorithm drove the robot to look for whatever was new.

   This asked a different question from standard optimisation: what happens when a chemistry robot is not told what to maximise? A curiosity algorithm controlled a formulation platform exploring self-propelling oil-in-water droplets, selecting experiments likely to reveal new behaviour. Against random exploration on the same budget, it observed far more diverse droplet dynamics and surfaced a specific sensitivity to temperature.

   Read the paper
6. Artificial intelligence exploration of unstable protocells leads to predictable properties and discovery of collective behavior PNAS2018L J Points, J W Taylor, J Grizou, K Donkers, L Cronin

   An automated platform explored unstable droplet systems and built models linking their physical properties to how they behave.

   The system selected and ran its own experiments, recording what happened through image recognition. From that data it built predictive models tying properties such as viscosity, surface tension and density to observed behaviours, including collective swarming. It turned a system usually treated as too unpredictable to study into one with describable rules.

   Read the paper
7. Self-calibrating BCIs: Ranking and recovery of mental targets without labels NeurIPS2026J Grizou, C De la Torre-Ortiz, T Ruotsalo

   A brain-computer interface that works out what the user is thinking of without ever being told the right answer.

   Conventional interfaces need a labelled calibration session before they can be used. This work recovers and ranks the user's mental target from unlabelled signals alone, removing that setup step. It is the same self-calibration idea Jonathan has developed across interfaces, applied here at a top machine learning venue.

   Read the paper
8. Interdisciplinary research in artificial intelligence: challenges and opportunities Frontiers in Big Data2020R Kusters, D Misevic, H Berry, A Cully, Y Le Cunff, L Dandoy, N Díaz-Rodríguez, M Ficher, J Grizou et al.

   What actually helps, and what gets in the way, when AI research crosses disciplines.

   Written from the experience of researchers working between AI and other fields, including chemistry and biology. The obstacles it describes, around shared vocabulary, incentives and tooling, are the same ones that show up when bringing automation into a working lab.

   Read the paper
9. Algorithm-driven robotic discovery of polyoxometalate-scaffolding metal-organic frameworks JACS2024D He, Y Jiang, M Guillén-Soler, Z Geary, L Vizcaíno-Anaya, D Salley et al.

   An algorithm steered a robot through the search for new metal-organic frameworks built on polyoxometalate scaffolds.

   Framework discovery depends on getting assembly and crystallisation right at the same time, which makes brute-force search expensive. Letting the search algorithm choose the next experiment focused effort on the productive regions of that space.

   Read the paper
10. Robotic stepwise synthesis of hetero-multinuclear metal oxo clusters as single-molecule magnets JACS2021T Minato, D Salley, N Mizuno, K Yamaguchi, L Cronin, K Suzuki

    A robot built metal oxo clusters one deliberate step at a time, producing single-molecule magnets.

    Stepwise construction gives control over exactly which metals end up where in the cluster, which is what determines the magnetic behaviour. Automating the sequence makes that control repeatable rather than dependent on a particular chemist's hands.

    Read the paper
11. Human versus robots in the discovery and crystallization of gigantic polyoxometalates Angewandte Chemie2017V Duros, J Grizou, W Xuan, Z Hosni, D-L Long, H N Miras, L Cronin

    A head-to-head comparison of human-guided and robot-guided strategies for finding and crystallising giant inorganic clusters.

    This explored how active machine learning can support discovery in a space where molecular self-assembly and crystallisation have to succeed together. Comparing the two strategies showed how structured experimental workflows expose the productive regions of a difficult chemical landscape.

    Read the paper
12. TELESIM: A modular and plug-and-play framework for robotic arm teleoperation using a digital twin ICRA2024F P Audonnet, J Grizou, A Hamilton, G Aragon-Camarasa

    A modular framework for driving a real robot arm through a digital twin of it.

    The operator works against a simulated copy of the arm, and the framework maps that onto the physical hardware. Being plug-and-play means it is not tied to one specific robot, which matters when a lab's hardware changes.

    Read the paper
13. Calibration-free BCI based control AAAI2014J Grizou, I Iturrate, L Montesano, P-Y Oudeyer, M Lopes

    Controlling a machine from brain signals with no calibration session first.

    Brain-computer interfaces normally need a training phase where the user produces known signals so the system can learn to read them. This work removed that step, letting the system infer the meaning of signals while being used. It is the foundation of the self-calibration line of work.

    Read the paper
14. Interactive learning from unlabeled instructions UAI2014J Grizou, I Iturrate, L Montesano, P-Y Oudeyer, M Lopes

    A robot learns both the task and how to interpret its teacher's signals, at the same time, with no dictionary given.

    Normally a robot is told in advance what a given human signal means. Here it works both out together, which is what allows a system to adapt to a new user without setup.

    Read the paper
15. Robotic modules for the programmable chemputation of molecules and materials ACS Central Science2023D Salley, J S Manzano, P J Kitson, L Cronin

    Standard robotic modules that make common chemical operations programmable across molecules, materials and formulations.

    Before a lab can benefit from AI it needs reliable digital interfaces to its experiments. This work is really an engineering philosophy: robust automation comes from modularity, repeatable interfaces and software-controlled workflows scientists will actually use. It bridges research prototypes and lab infrastructure that can scale past a single bespoke machine.

    Read the paper
16. A modular programmable inorganic cluster discovery robot for the discovery and synthesis of polyoxometalates ACS Central Science2020D S Salley, G A Keenan, D-L Long, N L Bell, L Cronin

    A modular robot platform built specifically for searching inorganic cluster chemistry.

    Cluster discovery depends on self-assembly and crystallisation happening under the right conditions, making brute-force search expensive and hard to reproduce. This platform combined high-throughput reaction execution with inline analysis. It shows that useful automation is not just about moving liquids: the platform has to match the scientific search problem.

    Read the paper
17. Intuition-enabled machine learning beats the competition when joint human-robot teams perform inorganic chemical experiments J. Chem. Inf. Model.2019V Duros, J Grizou, A Sharma, S H M Mehr, A Bubliauskas, P Frei et al.

    Human intuition combined with machine learning outperformed either working alone.

    This followed the human-versus-robot comparison by testing the middle ground: joint teams where a chemist's intuition, machine learning and automated experimentation all feed the same search. The combination beat the alternatives, which is the practical argument for human-in-the-loop automation rather than full replacement.

    Read the paper
18. Optimization of formulations using robotic experiments driven by machine learning DoE Cell Reports Physical Science2021L Cao, D Russo, K Felton, D Salley, A Sharma, G Keenan et al.

    Machine-learning-guided design of experiments, executed by robots, to optimise formulations.

    Formulation work involves many interacting ingredients, so the number of possible combinations grows fast. Using machine learning to choose which experiments to run keeps the number needed manageable, and the robot executes them consistently.

    Read the paper
19. Automated digital discovery and synthesis of CuO-based nanoparticle heterostructures for catalysis ACS Appl. Mater. Interfaces2025D Hervitz, Y Jiang, D Salley, M McNulty, P J Kitson, L Cronin

    Automated discovery and synthesis of copper-oxide nanoparticle heterostructures for use as catalysts.

    The platform searched combinations of synthesis conditions and made the resulting materials itself. Recording the process digitally is what makes a promising result reproducible later.

    Read the paper
20. AI-driven robotic crystal explorer for rapid polymorph identification Digital Discovery2026E C Lee, D Salley, A Sharma, L Cronin

    A robot that hunts for the different crystal forms a compound can take.

    The same molecule can crystallise into several polymorphs with different physical properties, which matters a great deal in pharmaceuticals. Automating the search covers more of the possible conditions than manual screening, and does it consistently.

    Read the paper
21. Exploiting task constraints for self-calibrated brain-machine interface control using error-related potentials PLOS ONE2015I Iturrate, J Grizou, J Omedes, P-Y Oudeyer, M Lopes, L Montesano

    Using what the task itself constrains to calibrate a brain-machine interface automatically.

    Error-related potentials are brain signals produced when someone sees a mistake. By combining those signals with the structure of the task, the system calibrates itself during normal use rather than in a separate setup session.

    Read the paper
22. Learning legible motion from human-robot interactions Int. J. of Social Robotics2017B Busch, J Grizou, M Lopes, F Stulp

    Teaching a robot to move in a way that makes its intention obvious to the person watching.

    A robot can reach a goal by many paths, and some make it much clearer what it is about to do. Learning that from real interactions makes collaboration safer and less ambiguous.

    Read the paper
23. Facilitating intention prediction for humans by optimizing robot motions IROS2015F Stulp, J Grizou, B Busch, M Lopes

    Optimising robot motion specifically so humans can predict what it will do next.

    Rather than optimising only for speed or efficiency, the motion is chosen to be readable. In a shared workspace, a person who can anticipate the robot works with it more comfortably.

    Read the paper
24. The evolution of active droplets in chemorobotic platforms ECAL / Artificial Life2017L J Points, J Grizou, J M Parrilla Gutierrez, J W Taylor, L Cronin

    How droplet systems change over successive generations when a robot runs the selection.

    A summary of what the chemorobotic droplet work showed about evolution in a purely chemical system. It sits alongside the Nature Communications and PNAS protocell papers as the artificial-life framing of the same platform.

    Read the paper
25. Poppy: open source 3D printed robot for experiments in developmental robotics ICDL-EpiRob2014M Lapeyre, P Rouanet, J Grizou, S N'Guyen, A Le Falher, F Depraetre et al.

    An open-source, 3D-printed humanoid robot built so other labs could reproduce and modify it.

    Poppy was designed to be fabricated on accessible hardware and openly shared, so experiments could be repeated elsewhere. It is an early example of the same principle behind the custom lab platforms: design for other people to build and change, not just for one demo.

    Read the paper