Lab Automation With a Low-Cost Robot Arm: Honest Limits

Start with the tolerance, not the arm

Split a bench workflow in two. There is metrology, where a number has to be right: aspirate 100 microlitres, read at 450 nm. And there is logistics, where an object has to end up somewhere: this plate into the reader, these tubes into the centrifuge. Metrology belongs to instruments. Logistics is what an arm does. A SO-100 does a surprising amount of the second and essentially none of the first.

What an SO-100 is, measured rather than advertised

The SO-100 and the newer SO-101 are printed arms built on Feetech STS3215 bus servos. The SO-ARM100 build repository prices one follower arm at USD 121.94 or EUR 124.30 and a leader-follower pair at EUR 226.30, before filament. The follower runs six STS3215 motors with 1/345 gearing per the LeRobot SO-101 guide; the leader mixes 1/191, 1/345 and 1/147 so a human can backdrive it.

Vendor pages quote payloads. The build repo does not, and that silence is honest. What you get instead is measured servo behaviour, from a bench test published as Testing of Feetech STS3215 Servomotor. Its torque figures are for the 30 kg.cm 12 V variant; the SO-100 follower runs the 7.4 V motor, rated at 16.5 kg.cm stall at 6 V, so read that column as a ceiling you will not reach.

The arm has six degrees of freedom including the gripper, so five for pose. Enough to lay a plate on a nest or drop a vial in a hole. Not enough to reach into a crowded deck at an arbitrary orientation while dodging a cable. Design the layout around the arm.

Job 1: pipetting. The answer is no

ISO 8655-2:2022 sets maximum permissible errors for air-displacement and positive-displacement pipettes complete with their selected tips and consumable parts, at 20 C, 50 percent relative humidity and 101.3 kPa. It scopes the instrument and the disposable, not the thing holding them. For scale, an Opentrons OT-2 lists 15 percent accuracy at 1 uL, its P300 single channel hits plus or minus 0.6 percent at 300 uL, and the machine weighs 48 kg and starts at USD 10,000.

• Force. The METTLER TOLEDO Rainin pipetting whitepaper puts tip ejection as the heaviest force in the cycle, four common pipettes spread across roughly 1 to 4 kgf. An SO-100 gripper does not make a coaxial 30 N push.
• Stiffness. A press fit needs the shaft coaxial as force builds. With 0.87 degrees of backlash, the wrist deflects first.
• Sealing. Too little tip loading force and the seal leaks, a volumetric error no camera will show you.
• Contamination. PLA's glass transition sits near 60 C, so a 121 C autoclave cycle is unavailable.
• Traceability. No calibration certificate and no path to one. Regulated workflows are out by construction, not performance.

Job 2: plate handling, with a fixture

Here standardisation rescues you. ANSI/SLAS 1-2004 fixes the microplate footprint at 127.76 mm by 85.48 mm, plus or minus 0.25 mm within 12.7 mm of the corners and plus or minus 0.5 mm along the side, with a flange corner radius of 3.18 mm. Every SBS plate has that outline, so one printed nest with generous lead-in chamfers provides the last millimetre instead of the arm.

The failure modes are specific. A plate gripped on its short edges skews if the jaws close asymmetrically, and drops if the gripper servo trips its overload protection mid-lift. A lidded plate is two objects that can separate. A full plate is not rigid, so sloshing moves the centre of mass: slow the trajectory rather than expecting the policy to compensate. Dropped objects usually trace back to gripper does not close.

Job 3: sample sorting, the sweet spot

Sorting is where a learned policy beats a script and the arm's weaknesses stop mattering. Rack holes give a millimetre or two, the objects are light, there is no force spec, and the decision is visual, which is what a vision-language-action model is for.

Published work brackets this. Burger et al., A mobile robotic chemist (Nature 583, 2020) ran 688 experiments over eight days and framed the contribution as automating the researcher rather than the instruments. At the other end, Conrad et al. (Advanced Intelligent Systems, 2025) moved specimens from a magazine to a pallet with a desktop Dobot MG400, reporting 100 percent correct placement at 8.9 s per transfer, 2.8 times faster than the human demonstration.

Teach the motion instead of programming it

Inverse kinematics plus waypoints works when the world is fixed, and a bench is not: racks shift, tube heights vary, someone leaves a marker pen on the deck. The imitation learning route is to drive the arm through the task a few dozen times and train a policy on what the cameras and joints saw. The LIRA paper (Communications Chemistry, 2025) states the problem plainly: manipulation in most self-driving lab workflows is open loop, with no real-time error detection. Their closed-loop vision fix reached 1.02 mm mean translation error using 75 mm ArUco markers, at about 5 s per calibration against roughly 53 s tactile.

# LeRobot plus the Feetech SDK the STS3215 bus needs
pip install -e ".[feetech]"

# Which serial port is the arm on? Unplug when prompted.
lerobot-find-port

# One-time: write ids and baudrate into each servo's EEPROM,
# gripper first, one motor connected at a time.
lerobot-setup-motors --robot.type=so101_follower --robot.port=/dev/ttyACM0

# Calibration: mid-range pose, then sweep every joint end to end.
lerobot-calibrate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=bench_arm

1. 1
   Fix the scene before recording anything

   Bolt down the nest, the rack and the cameras, and mark their outlines. A policy trained on a moving camera learns nothing transferable. Two views work well: overhead for the tray, wrist for the grasp.

   bash

   lerobot-teleoperate \
       --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=bench_arm \
       --robot.cameras="{ top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30} }" \
       --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=bench_leader \
       --display_data=true

2. 2
   Record, varying only what should vary

   Vary tube position, roughly 10 episodes per region. Do not vary lighting, camera pose or grasp style yet. Defaults: episode_time_s 60, reset_time_s 60, num_episodes 50.

   bash

   lerobot-record \
       --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=bench_arm \
       --robot.cameras="{ top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30} }" \
       --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=bench_leader \
       --dataset.repo_id=${HF_USER}/tube-sort \
       --dataset.num_episodes=60 \
       --dataset.episode_time_s=30 \
       --dataset.reset_time_s=15 \
       --dataset.single_task="Put the capped tube into the empty rack hole" \
       --display_data=true

3. 3
   Replay before you train

   Replay drives the recorded joint targets back at the arm with no policy involved. If it misses the rack hole, the problem is mechanical, and training will only launder it.

   bash

   lerobot-replay \
       --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=bench_arm \
       --dataset.repo_id=${HF_USER}/tube-sort --dataset.episode=0

4. 4
   Train the cheap model first

   ACT trains from scratch at roughly 80 M parameters on a 24 GB card and costs 20 ms per action step. Save the 3 B models for when ACT has plateaued.

   bash

   lerobot-train \
     --dataset.repo_id=${HF_USER}/tube-sort \
     --policy.type=act \
     --output_dir=outputs/train/act_tube_sort \
     --job_name=act_tube_sort \
     --policy.device=cuda

5. 5
   Roll out and count failures by cause

   Log why each attempt failed: missed the tube, dropped it, missed the hole, hit the rack. Four buckets, four fixes. In current LeRobot, --strategy.type=base runs autonomously without recording.

   bash

   lerobot-rollout \
     --strategy.type=base \
     --policy.path=${HF_USER}/act_tube_sort \
     --robot.type=so100_follower --robot.port=/dev/ttyACM0 \
     --task="Put the capped tube into the empty rack hole" \
     --duration=120

Which policy for a bench task

Five models are trainable here, and for lab logistics the choice is latency and episode count. ACT is the default for a fixed bench; SmolVLA is the alternative when you want language conditioning or cannot record 50 episodes. The full comparison and ACT against SmolVLA have the rest.

Where the policy has to run

A bench arm has an unusual advantage: nobody minds if the plate takes four seconds instead of one, which makes remote inference viable in a way it is not for reactive work. The constraint is that the control loop already costs 20 to 485 ms per action step, and public-internet round trips on top turn a working policy into a hesitant one. Slow pick and place is fine on a cloud pod; anything reacting to contact needs inference next to the servos. See inference latency and action chunking.

What this does not fix

Useful for logistics is not the same as useful. Several things a lab needs sit outside what any 110 EUR arm delivers, and pretending otherwise is how automation projects die at month four.

• Error recovery. A trained policy shares the open-loop flaw the LIRA authors describe: it continues after dropping the tube, because nothing told it. Their inspection module reached 97.9 percent detection accuracy as a separate build.
• Device integration. Moving a plate is one command in a workflow that also talks to a reader, a shaker and a LIMS. That layer is what SiLA 2 and PyLabRobot exist for. The arm is a peripheral, not the system.
• Throughput. One arm at a 10 second transfer is 360 moves an hour at best, with servo heating in the way.
• Validated volume. Without a balance or a reader in the loop, a pipetting policy's error is unobservable.
• Generalisation. A policy is bound to the camera poses and lighting it saw, so rebuilding the bench means rebuilding the dataset: policy only works in one setup.
• Anything regulated. No calibration certificate, no qualification path, no autoclave. A hard boundary, not a roadmap item.

A realistic first project

Build a tube handoff. Not a workflow, not a pipetting cell. One arm, one tray, one rack, one decision. Published low-cost work sits at this scale: Logan, Dudash and Negron (arXiv, March 2026) report a modified 5-DOF arm with a liquid handling end effector reaching mean positional error below 1 mm for colony picking, with segmentation at 0.537 IoU. Sub-millimetre placement is achievable at this price. Reliable perception is the harder half.

1. 1
   Week 1: fixtures and replay, no learning

   Print a nest and rack holder against the SLAS dimensions with 3 mm or larger chamfers, bolt them and the cameras down, run the calibration sweep, then record five throwaway episodes and replay them. Nothing downstream rescues a nest the arm cannot hit.

2. 2
   Week 2: record 60 episodes in one sitting

   One person, one session, one grasp style, tube position varied across the tray. At 30 s per episode plus 15 s reset, roughly 45 minutes.

3. 3
   Week 2: train ACT and measure

   Train ACT, run 30 rollouts and tally the four failure buckets. A number you wrote down beats a video you remember.

4. 4
   Week 3: add the balance

   Put a scale under the destination and log mass per handoff. That independent check is the difference between a demo and an instrument.

5. 5
   Week 4: decide honestly

   Above roughly 95 percent success on a task you repeat, keep it and add error detection. At 80 percent it is a teaching tool and the workflow still needs a human.

Before committing, the model arena compares 85 VLA models across 332 benchmark results, and the three ways to start without hardware lets you drive a real arm first. Background reading: the complete SO-100 setup and training guide, the overview of vision-language-action models, and how to collect high-quality VLA training data.

For the state of the art, LabVLA (arXiv, June 2026) grounds a VLA in laboratory protocols on a Qwen3-VL-4B-Instruct backbone, evaluated on the LabUtopia benchmark. And ORGANA remains the clearest statement of the assistive framing, with a chemist in the loop and 80.3 percent average time saved. Note what ORGANA automates. Not the measurement. The hands.