assemblab / Formulation SDL Atlas

Liquids · pastes · powders · objects

A formulation lab is a physical choreography.

Ingredients of very different rheology are dosed into one premix cup, carried to a mixer, processed and characterized. This atlas follows that cup through the seven questions a team must answer to automate it, using the data of our review and a live digital twin of a formulation cell.

178references
7comparison tables
20literature scenarios
4robot arms in the twin
Graphical abstract: a formulation SDL dosing liquids, pastes and powders into a premix container that robots move to a mixer and characterizers, mirrored by a digital twin and directed by AI agents.
Graphical abstract of the review.
Loading twin…

Real UR e-Series kinematics and meshes; every motion is an RRT-Connect plan checked against the device boxes before it plays.

The cell

One cup, four stations

The twin on the right is the Self-Formulating Lab cell: an analytical balance, a CNC dispensing gantry, a planetary mixer and a characterizer around a collaborative arm. It is the same cell as Fig. 1 of the review, reduced to the stations a premix cup visits.

Figure 1: a physical formulation cell with a CNC gantry, balance, SCARA elevator, characterizers and an arm loading a mixer, beside its digital twin.
Fig. 1. A formulation SDL and its digital twin.

Scroll on. Each step below moves the camera to the part of the cell it is about.

1 Dose

How is each ingredient dosed?

Viscosity, yield stress and powder flowability decide the mechanism. Drag the marker along any scale to see which technologies the review recommends there (Table 1).

2 Move

How does the premix cup move?

Five robot classes carry containers between stations, ordered by axes, reach and cost. Price tracks the class and the ecosystem more than the repeatability (Table 2). Select a point for details.

3 Buy

What can be bought?

Commercial dispensers by material class and integrated liquid-handling platforms (Tables 3 and 4). Bars show list prices on a log scale; hollow tags are quote-only, which is common where handling is hardest.

4 Extend

How do you go beyond the catalogue?

Low-cost retrofits change one layer of a purchased platform and unlock a capability its maker never specified (Table 5).

5 Rehearse

Will the plan work?

A twin rehearses a workflow before material is committed. Fidelity rises in four levels, and so does compute cost. The twin here works at level 2: kinematics and collision checking, shown by the boxes it plans against (Table 6).

Figure 7: the digital twin as a working environment, with MoveIt 2 motion planning, a GPU simulator, a DEM/MPM material solver and rehearsal of AI-proposed plans.
Fig. 7. The digital twin as a working environment.

6 Orchestrate

How do the devices run as one?

Orchestration software sits between decisions and devices: it schedules work, holds the resource model, drives instruments through their interfaces and records what happened. The review compares the functions of eight open-source systems.

Figure 8: architecture of a formulation SDL with execution and equipment hierarchies, functional layers of orchestration software and functions of eight open-source systems.
Fig. 8. Architecture of a formulation SDL.

7 Decide

What should we make next?

Data at the bottom, models that choose the next experiment in the middle, AI agents on top. Each agent role gets an autonomy level and a human checkpoint (Table 7).

Figure 9: the decision ladder from data to models to AI agents, connected to robots by software and data pipelines.
Fig. 9. The decision ladder.

Fig. 3 of the review

Twenty ways labs sense, then act

Each study is titled by its sense–act sequence. Read across a row and a pattern appears: methods that measure the material before or while dosing it generalize best.

    What we are building

    From a review to a lab that designs itself

    The review is the knowledge. The wiki keeps it current. Two agents put it to work: one watches vendors, the other designs a lab for a user's budget, materials and room.

    1. Complete

      Review

      Self-Driving Laboratories for Formulation. Eight authors, 178 references, seven tables.

      Cite
    2. Updated to V3

      Wiki

      Living directory of materials, strategies, hardware, prices and tools, now rebuilt from the final manuscript.

      Open the wiki
    3. Designed

      Library agent

      Watches vendor releases and proposes catalogue changes with source and date. A person approves each one.

      Spec
    4. Prototype

      SFL agent + Lab Builder

      Talks with the user, proposes a cell, and shows it in the twin above with collision-checked motions. In the Lab Builder, users drag commercial devices into a 3D workstation and the cart prices it live.

      Open the Lab Simulator

    Plan a cell

    A preview of the SFL agent's selection step. It is rule-based and uses list prices from the review. To lay the cell out yourself, open the assemblab Lab Simulator.

    Ingredients you will handle

    Cite the review

    The paper behind the atlas

    Manuscript in preparation (2026). The tables on this page are the review's Tables 1–7; numbered links go to its reference list in the wiki.

    Related work

    the self-building lab planner

    the self-building lab planner (Bai, Zhao et al., arXiv 2610.23826) build the layout, twin, robot skills and workflows a lab needs for a scientific goal. The Self-Formulating Lab works one step earlier, choosing what to buy and how to lay it out for a formulation budget, so its output could be a starting point for a self-building lab.