ROBOTIC LEARNING IN THE WET LABOCTOBER 2026

LabApprentice:
Self-Improving Wet-Lab Robotics Harness

LabApprentice explores how a coding agent can accumulate practical tools through physical experiments. Programs written for pouring, heating and instrument control become available to later runs.

Explore the experiments

Experiments

05 tasks / 18 recorded runs

Bringing the tools together

A hydrogel experiment using accumulated tools

Make a PEDOT:PSS hydrogel in two molds and measure each gel with the potentiostat, starting with the store from Pour, Heat, Split and Calibrate.

Hydrogel-1 combines the earlier tasks in one workflow.

Hydrogel experiment film poster

Loading video…

Hydrogel · 4:29Playback speed and camera views are labelled in the film.

MOMENTS TO WATCH

The arm turns the STIR knob with the program from Heat.

A lid missing from mold 2 is found on the bench and placed again.

The potentiostat measures the gel with the program from Calibrate.

In these runsHydrogel-1 reuses the knob program from Heat and the potentiostat program from Calibrate. During the run, a missing lid is found on the bench and placed again.

Keeping the programs

From one run to the next

Each run begins in a fresh agent session with a written procedure and run-specific parameters. The agent reads a persistent skill store, then reuses or writes programs to carry out the experiment. The store holds tools, written skills, notes, and dated bench positions and calibrations.

The starting harness supplies basic motion, perception and instrument tools. During a run, the agent adds missing capabilities to the store. The store is copied before and after each run; people also correct stored facts between runs, and the agent can record their answers.

  1. 01 / READ

    Procedure and prior tools

    A new session reads the task, its parameters and the store’s notes before acting.

  2. 02 / EXECUTE

    Code meets the bench

    Reused or newly written programs operate the instruments. Every arm motion passes through a server with fixed motion checks.

  3. 03 / VERIFY

    Claim, then check

    For each evaluated step, the agent records its belief before reading the instrument, then logs the reading and verdict against the criterion.

  4. 04 / KEEP

    Pass on usable artifacts

    Programs, instructions and practical observations stay in the store for the next session.

Checks come from the instruments

The balance reports mass over RS-232. Cameras read the hot plate’s displays and record spills. The potentiostat measures impedance through its vendor toolkit. The arm turns the hot plate’s physical knobs.

Supervised physical experiments

A person holds the emergency stop and responds to logged requests, including vessel exchanges and cable changes. People label the criteria from the records after each run. The harness’s programs and skills remain fixed within each task.

Pour

Mass-targeted pouring

pour_target.py

Heat

Turning instrument knobs

Knob program

Split

Grasp and pour corrections

Stop 0.8 g early

Calibrate

Potentiostat control

Instrument program
HYDROGEL-1

Tools reused in Hydrogel-1

Heat → STIR-knob control. Calibrate → potentiostat measurement.

A COMPARISON WITH AN EMPTY STORE

What happens when the saved tools are absent?

Runs ending in -C repeat the task’s second variant with a fresh, empty store, after the runs using the accumulated store. The harness, procedure, person and bench are held the same. Across the repeated runs, targets, vessel orientations or instrument positions change.

  • Pour-CWrites the pouring program again.
  • Heat-CWorks out a tilted grasp from scratch.
  • Split-CA beaker slip is visible in the camera frames.
  • Calibrate-CDeforms a mold without the stored grip rule.

Measurements

Recorded experiments · October 2026

Agent-and-robot time combines reasoning and programming with robot execution. Instrument readings, procedure waits and waits for the person are accounted for separately.

First run, last run with the store, and empty-store control Time in minutes:seconds · shorter is faster

Pour

59% less time, first to last
First (Pour-1)
21:22
Last (Pour-3)
8:40
Control (C)
23:12

Heat

22% less time, first to last
First (Heat-1)
25:07
Last (Heat-2)
19:37
Control (C)
25:22

Split

54% less time, first to last
First (Split-1)
48:23
Last (Split-5)
22:17
Control (C)
78:46

Calibrate

24% less time, first to last
First (Calibrate-1)
20:14
Last (Calibrate-3)
15:27
Control (stopped)
34:06

Each bar is one run. Targets, positions or vessel orientations changed between runs; each control repeats the second variant with an empty store. The Calibrate control was stopped after a mold cracked.

Hydrogel-1

The combined workflow ran 13 of the 14 stored tools from the earlier tasks. All 8 evaluation criteria were met in this run.

After curing at 60 °C for 2 hours, the potentiostat measured the two gels. Conductivity was estimated from resistance and each mold’s calibrated cell constant.

Measurements from the two molds
MoldDispensed mass (g)Resistance (Ω)Conductivity (S/m)
15.26142.30.71
25.62115.81.12

Data from the LabApprentice manuscript.

Scope of this study

These are preliminary wet-lab experiments. The Hydrogel demonstration consists of one recorded run. The examples document specific instances of program reuse and physical execution; broader reliability and generalization require further evaluation.

Reading the recordings

Pour, Split and Calibrate use snapshots taken at the agent’s steps. Heat and Hydrogel also include timelapse footage. On-screen checks, instrument readings and end cards provide context for each run.

Film timestamps refer to the edited videos. Agent-and-robot time combines reasoning and programming with robot execution. Instrument time, procedure waits and human waits are accounted for separately. The recordings also show a wall-clock counter.