AEGIS: Assay-Aware Protocol Validation and Runtime Monitoring for Open-Source Liquid Handling Robots

Arvind Ramanathan; Ian Foster; Priyanka V.Setty; Rick Stevens

REVIEW 3 major objections 4 minor 15 references

AEGIS catches silent pipetting failures and assay-level protocol errors on open-source liquid handlers with a camera and an LLM rule checker.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 22:44 UTC pith:D5PL2EB3

load-bearing objection Honest, well-scoped proof-of-concept for a useful integration, but the deployment-faithful numbers rest on an untested assumption about the calibration wells, and the benchmark is small; worth refereeing with requests for fixes. the 3 major comments →

arxiv 2607.15620 v1 pith:D5PL2EB3 submitted 2026-07-17 cs.RO cs.AI

AEGIS: Assay-Aware Protocol Validation and Runtime Monitoring for Open-Source Liquid Handling Robots

Priyanka V.Setty , Arvind Ramanathan , Ian Foster , Rick Stevens This is my paper

classification cs.RO cs.AI

keywords liquid handling robotOpentrons OT-2protocol validationruntime monitoringanomaly detectionPCA world modelvision-language modelself-driving laboratory

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a two-layer guardian can close two silent-failure gaps that corrupt data in self-driving laboratories built on open-source liquid handlers like the Opentrons OT-2: protocols that compile but violate assay biology, and physical execution failures that no log will flag. Layer 1 pairs a curated 22-rule database with a single LLM pass over the protocol's Python and reaches F1 0.97 on a 24-protocol benchmark, never missing an injected bug. Layer 2 fits a PCA world model to YOLO-cropped four-frame pipette trajectories and, under a leave-one-plate-out evaluation, detects failure wells at AP 0.89 and F1 0.71, with a cascade that routes ambiguous wells to a vision-language model that alone can pause the robot. On a physical OT-2 the system catches planted no-tip failures deterministically and partial dispenses on coloured dyes, and it holds VLM cost near $1.63 per plate. If correct, open-source labs can retrofit runtime and pre-flight monitoring with commodity hardware, and the paper's honest reporting of water and small-pipette limits marks where the approach currently stops.

Core claim

The central claim is that AEGIS is the first system to unify pre-flight assay-aware validation with runtime visual monitoring for an open-source liquid handler. Layer 1 (rules + LLM) reaches adjusted F1 0.97 on 24 protocols with 100% recall of injected bugs; Layer 2 (PCA world model on YOLO-cropped four-frame trajectories) reaches AP 0.89 / F1 0.71 under leave-one-plate-out, with a cascade where a VLM self-vote is the sole pause arbiter and per-run calibration on the first six wells anchors the threshold. Cost is held near $1.63 per 96-well plate. The paper also reports honest limits: water is a principled front-view boundary, p20 sits at F1 0.47, and air bubble is the weakest mode.

What carries the argument

The carrying mechanism is a two-layer cascade. Layer 1 is a hybrid validator: a machine-readable 22-rule database, the OT-2 protocol source, and an in-context API reference feed one LLM call that classifies the assay, cites violated rules, and passes or fails the protocol. Layer 2 is a train-on-normal trajectory world model: YOLO tip crops from four keyframes (ASPIRATE, TRANSIT, DISPENSE, POST-DISPENSE) form a 9,216-dimensional vector, and a 15-component PCA reconstruction error is the anomaly score. A per-run calibration band (median ± 0.7·MAD on the first six wells) routes wells to SILENT_PASS, BLANK_FLOOR, or GRAY_ZONE; only the latter two reach a three-lens VLM self-vote that alone can p

Load-bearing premise

The system assumes the first six wells of every run are physically normal, because it sets its detection threshold from those wells; if a run's opening wells are abnormal, the threshold is mis-anchored and failures can pass silently.

What would settle it

Plant a failure in one of the first six wells of a plate (e.g., a no-tip or partial dispense), then plant the same failure later in the same run, and check whether the later failure is silently passed because the calibration band widened; a system with mis-anchored calibration would miss it, while the paper's reported numbers presume normal opening wells.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Any OT-2 lab can retrofit both pre-flight and runtime monitoring with a webcam, a laptop, and the existing Opentrons API, with no robot-side hardware changes.
Assay-aware validation does not depend on paid frontier APIs: the free open-weight model matches the best proprietary backend at F1 0.97, so the rule database, not model scale, supplies the precision.
The cascade's cost dial means operators choose between $1.63/plate (default) and $10.33/plate (always-VLM), an explicit safety-vs-cost trade rather than a baked-in threshold.
Failures are caught with complementary detectors: no-tip deterministically via BLANK_FLOOR, partial dispense best by the VLM, no-aspirate best by the PCA transit-frame signal; routing everything to one detector loses coverage.
The released benchmark (24 protocols, 22 rules, 128 labelled trajectories) is the first public corpus of liquid-handling failures with assay-level rule annotations, giving later systems a shared testbed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

A testable extension follows directly from the calibration design: if a run's first six wells are not physically normal, the per-run threshold mis-anchors; replacing the single-run calibration with cross-run normal priors accumulated across sessions would remove this vulnerability.
The triage-and-validate cascade pattern is domain-general: any setting with a cheap but imperfect anomaly detector and a fallible but capable LLM gate (e.g., drone inspection, industrial quality control) could adopt the same 'cheap model routes, expensive model decides' architecture.
The p20 and water results point to a concrete hardware path: a tip-proximal camera or a top-down post-run plate view (the paper's prospective Layer 3) would attack the resolution limit and measure dispensed volume independently of reagent transparency, potentially lifting the weakest failure modes.
If the reported LOPO numbers transfer across OT-2 deployments, the practical bottleneck shifts from accuracy to the ~9.5 s monitor-side latency; a faster triage model or signal-driven frame flush would be the highest-leverage improvement, as the paper itself notes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Honest, well-scoped proof-of-concept for a useful integration, but the deployment-faithful numbers rest on an untested assumption about the calibration wells, and the benchmark is small; worth refereeing with requests for fixes.

read the letter

First, the thing you should know: this is an honest proof-of-concept of a genuinely useful integration, and the authors have done a better-than-usual job of reporting their own failure modes. But the two headline numbers — Layer 1 adjusted F1 0.97 and Layer 2 deployment-faithful F1 0.71 — both have soft spots that are addressable but need airing.

What's actually new: AEGIS is the first system I know of that combines assay-aware pre-flight protocol validation with runtime visual monitoring on the OT-2. The 22-rule database, the 24-protocol benchmark, and the trajectory dataset are new artifacts; the cascade design pattern (PCA triage, VLM gate, neither can pause alone) is sensible and the cost/latency analysis is useful. The live demo on a physical OT-2 with five replicates per condition, including honest reporting of water being a principled limit, gives real evidence.

The reporting is genuinely candid: negative results (air bubble 1/6, p20 F1 0.47, water 0/5), Wilson intervals at small N, a limitations section that names the calibration-cluster concern. That's rare and should be credited.

The soft spots, in proportion. The largest is the calibration assumption: the cascade anchors its silent-pass band to the first six wells, and the paper assumes these are physically normal. The stress-test note is right — no experiment plants a failure in the calibration window, and the offline LOPO always uses a known-normal well, so the deployment-faithful claim is conditional on an assumption that is plausible but untested. This is not fatal to the concept, but it's the first thing I'd ask the authors to stress-test.

Second, the LOPO isn't fully leakage-free: the threshold is calibrated on the held-out plate's own first normal well, and the PCA component count k=15 is selected on the full normal set, not re-selected per fold. These are mild leaks; a fully nested selection and a strict holdout for threshold calibration would tighten the numbers.

Third, the Layer 1 headline is an adjusted F1 after relabeling two "correct" protocols as buggy. The authors say they report both strict and adjusted, but the results section only shows adjusted. Presenting both would be more fair, especially since the strict numbers would be lower (though still strong).

Fourth, the open-source assets are placeholders (GitHub URL and Zenodo DOI to be added), so reproducibility is not yet verifiable. And the benchmark is small — 24 protocols, 62 scored trajectories — which means the intervals are wide.

None of this sinks the paper. The central proof-of-concept holds up: a cheap retrofitted guardian can catch important silent failures. The right venue is peer review with a request for code/data release, nested hyperparameter selection, a stress-test of the calibration assumption, and strict/adjusted Layer 1 scores reported separately.

Referee Report

3 major / 4 minor

Summary. AEGIS is a two-layer guardian for the Opentrons OT-2. Layer 1 is a pre-flight validator that combines a curated 22-rule assay database with an LLM reading OT-2 Python protocols; it is evaluated on 24 protocols (11 correct, 13 with an injected single-bug) across five assay families, reaching adjusted F1 0.97 and 100% injected-bug recall. Layer 2 is a runtime visual monitor that crops four-frame pipette trajectories with YOLO, projects them via PCA, and routes ambiguous wells to a VLM gate. Under a leave-one-plate-out protocol it reports AP 0.887, AUROC 0.804, and operating-point F1 0.708 on the p1000 channel, with a live demonstration on a physical OT-2 catching planted no-tip, partial-dispense, and no-aspirate failures. The paper includes honest reporting of negative results, confidence intervals, a transparent-reagent resolution limit, and cost estimates ($1.63/plate for the cascade).

Significance. If the reported numbers hold, AEGIS would be a practical, low-cost retrofit for the widely used OT-2, addressing two silent-failure modes that currently corrupt self-driving-lab data. The paper's strengths are its honesty: it reports negative results, small-N confidence intervals, the transparent-water limit, and the deliberate choice of a simple PCA baseline. The two-layer cascade design (cheap PCA triage, expensive VLM validator, neither allowed to act alone) is a transferable pattern, and the open release of benchmark, rules, weights, and reproduction scripts is valuable. However, the central Layer-2 evaluation rests on an internally inconsistent LOPO test-set description, a hyperparameter-selection procedure that is not fully leakage-free, and an unvalidated per-run calibration assumption. These issues must be resolved before the deployment-faithful claim is fully supported.

major comments (3)

[§2.3, §3.2.2, Fig. 6] The LOPO evaluation's test-set composition is internally inconsistent. The corpus is 62 scored trajectories: 26 physically-normal and 36 planted failures. §3.2.2 reports '6 false positives on the 10 held-out normal wells' and 'failure prevalence 0.58'. If only 10 normals are scored (because one normal per plate is reserved for threshold calibration), the scored set has 36+10=46 samples and the failure prevalence is 36/46=0.78, not 0.58. If all 26 normals are scored, prevalence is 0.58 but the '10 held-out normal wells' statement is wrong. Since AP, AUROC, and the operating point are all prevalence-dependent, please state exactly which wells are scored per fold and recompute the metrics and baseline accordingly.
[§2.5, §3.2.1] The 'leakage-free' claim is overstated because k=15 is selected once on the full set of 26 normal trajectories before the LOPO folds are run. The held-out plate's normal trajectories therefore contribute to the choice of model complexity, even though the PCA is refit per fold. With only 26 normals, this can inflate the reported transfer performance. Please either select k inside each LOPO fold (nested LOPO) or report the sensitivity of the LOPO results to k over the swept range. The current presentation does not justify the 'leakage-free' label.
[§2.6, §3.4 Limitations] The deployment-faithful F1 claim rests on the per-run calibration assumption that the first six wells of every run are physically normal, and the LOPO evaluation calibrates the threshold from the held-out plate's own first normal well. No experiment or simulation tests what happens when this assumption fails, even though the paper explicitly concedes the median±0.7·MAD band is 'robust to a single anomalous calibration well but not to a cluster of mis-behaving calibration wells' (Limitations). Because silent failures are the motivating scenario, an early-run anomaly would mis-anchor the band and could silently pass later failures. Please add a sensitivity analysis (e.g., inject one/two anomalous calibration wells and report downstream recall) or substantially soften the 'deployment-faithful' wording.

minor comments (4)

[Abstract; §1.3; §3.1] The abstract and conclusion report the Layer 1 F1 as if it were a general semantic-error detection result, while §1.3 correctly scopes it as a rule-coverage benchmark whose bug labels are generated from the same 22-rule database. Please carry this qualifier into the abstract and conclusion to avoid overstatement.
[§2.3, §3.2.2] The exclusion of the two NO_TIP samples from the offline quantitative evaluation is transparent, but the abstract's 'runtime visual monitoring' claim should make clear that the PCA-based AP/F1 numbers cover only the 36 non-no-tip failures; no-tip is handled by the separate BLANK_FLOOR rule and reported only in the live demo.
[Data availability] The paper promises an MIT-licensed release, but the GitHub URL is a placeholder ('to be added on release') and the Zenodo DOI is 'to be assigned upon acceptance'. For a reproducibility-focused paper, include the actual release identifiers or a stable repository at the time of submission.
[§2.6] The silent-pass band width factor (0.7·MAD) and calibration window size (n=6) are fixed hyperparameters, but no sensitivity analysis is reported. The paper notes in §3.4.0.4 that the cascade redesign changed several factors jointly; a small ablation of the band width and window size would strengthen the claim that the operating point is not tuned to the specific live protocol.

Circularity Check

1 steps flagged

Headline Layer-1 F1 is partly self-confirming: two benchmark labels are reset to 'buggy' because the evaluated models flagged them; Layer 2 is an honest train-on-normal evaluation with a stated but unvalidated calibration assumption.

specific steps

self definitional [Section 2.1 (protocol benchmark and failure injection); headline number in Abstract and Section 3.1]
"Two of the eleven nominally-correct tutorial protocols in fact contain a genuine tip-reuse bug; we therefore report both a strict score against the original labels and an adjusted score that relabels these two as buggy, crediting the models that flag them."

The adjusted ground truth for two of the 24 benchmark protocols is defined by the very model outputs it is then used to score: the protocols are relabeled 'buggy' because the evaluated models flagged them, so those flags are counted as true positives rather than false positives. The abstract's headline 'adjusted F1 of 0.97' therefore contains a component that is constructed from the model's own predictions, not from an independent label. The paper is transparent about the adjustment and mentions a strict score, but the main-text results are reported under the adjusted scoring, so the headline number is partly self-confirming. This does not affect the Layer 2 PCA/VLM evaluation, which uses planted physical-failure ground truth and is not circular.

full rationale

The only concrete circularity I can exhibit with a quote is the Layer-1 adjusted-label step: two benchmark protocols are relabeled as buggy because the evaluated models flagged them, and the headline F1 is computed under that adjusted scoring. That is a genuine, if narrow, evaluation circularity affecting the central Layer-1 claim. Layer 2 is structurally different and not circular: the PCA world model is fit on normal trajectories only, the anomaly score is reconstruction error, and the LOPO evaluation scores held-out plates against planted physical-failure labels. The per-run calibration on the first six wells is an assumption, not a circular derivation; the paper openly states its failure mode ('robust to a single anomalous calibration well but not to a cluster of mis-behaving calibration wells'), and no experiment validates it, but that is a robustness limitation rather than a reduction of the prediction to its inputs. Self-citations (PRISM, Academy) are contextual and not load-bearing for the main claims. The Layer-1 rule database and injected bugs share the same rule vocabulary, but the paper explicitly frames Layer 1 as a rule-coverage benchmark, so that overlap is an acknowledged design choice rather than hidden circularity. Overall, the central Layer-2 result has independent content, and the Layer-1 circularity is limited to the adjusted-label adjustment; hence a moderate score of 3.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The evaluation rests on small hand-curated datasets (24 protocols, 62 scored trajectories), a hand-chosen calibration band, and the domain assumption that opening wells are normal. No new physical entities or mechanisms are proposed; AEGIS is an integration of existing detectors with a domain-specific rule set.

free parameters (4)

PCA component count k = 15 (97.7% variance)
Selected once on the full 26-normal trajectory set before LOPO (Sections 2.5, 3.2.1); the k-sweep in the SI uses the same corpus to justify the knee, so the reported AP/F1 are tuned on all data.
Silent-pass calibration band width factor = 0.7 * MAD around median
Section 2.6 defines the band as median +/- 0.7*MAD; the 0.7 multiplier and the n=6 calibration window are hand-chosen and directly control which wells are routed to the VLM vs silently passed.
Per-plate reconstruction-error threshold = Calibrated from each plate's first normal well
In LOPO the threshold is set on the held-out plate's own first normal well (Section 2.3), using a test-plate sample to set the operating point; in live runs it is anchored to the first six wells.
Calibration window size = n=6
The cascade calibrates on the first six wells (Section 2.6); changing n changes the silent-pass band and the false-positive/coverage tradeoff.

axioms (5)

domain assumption The first six wells of a protocol are physically normal standards/controls
Layer 2 anchors its silent-pass band to these wells (Section 2.6); if a run's opening wells are abnormal, the cascade is miscalibrated.
domain assumption A front-view webcam at 1 fps with 48x48 tip crops contains enough geometric signal to distinguish normal from failed p1000 pipetting
The entire PCA/VLM pipeline assumes the trajectory images carry the failure signal; the paper's own p20 and water results show this degrades sharply at smaller tips and transparent fluids.
domain assumption YOLOv8-nano fine-tuned on 69 hand-labelled frames generalizes sufficiently that crop variability does not dominate PCA reconstruction error
The detector reaches mAP50 0.995 on a held-out split (Section 2.4), but the full evaluation depends on its out-of-distribution behavior.
domain assumption The four keyframes (ASPIRATE, TRANSIT, DISPENSE, POST-DISPENSE) capture failure-relevant dynamics
The trajectory representation is fixed by these protocol-level keypoints (Section 2.3); failures that appear only between frames would be missed.
standard math PCA reconstruction error is a valid anomaly signal for normal pipetting trajectories
Reconstruction error ||x - x_hat||^2 is used as the anomaly score (Section 2.5), a standard train-on-normal anomaly-detection assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 21089 in / 16783 out tokens · 157370 ms · 2026-08-01T22:44:28.319949+00:00 · methodology

0 comments

read the original abstract

Self-driving laboratories increasingly rely on low-cost liquid handlers such as the Opentrons OT-2, which ship without the pressure-based aspiration monitoring of Hamilton or Tecan systems and are typically run open-loop. Two failure modes go undetected: protocols that are syntactically valid but violate assay-specific invariants (e.g., tip reuse between a PCR template and a no-template control), and physical execution failures (partial dispense, air bubbles, missing tips) at runtime. We present AEGIS, a two-layer guardian for both. Layer 1 pairs a curated machine-readable assay rule database with an LLM that reasons over OT-2 Python code, reaching an adjusted F1 of 0.97 on a 24-protocol benchmark across five assay families and beating rules-only and LLM-only ablations across five backends; a free open-weight model ties the best proprietary one, so no paid API is required. Layer 2 fits a PCA world model to YOLO-cropped four-frame pipette trajectories; under a leakage-free leave-one-plate-out evaluation it reaches average precision 0.89 and operating-point F1 0.71 (AUROC 0.80), a deployment-faithful number that matches the live demonstration, and we characterize the small-pipette (p20) resolution limit (F1 0.47). A live demonstration on a physical OT-2 (five replicates per condition) catches planted no-tip failures deterministically and partial dispense on coloured dyes, with an always-VLM self-vote gate lifting partial-dispense recall to 5/5; transparent water is a principled limit of any front-view-only monitor, which AEGIS surfaces as low-confidence VLM reasoning rather than a wrong verdict. Cascade triage holds VLM cost near $1.63 per plate versus $10.33 for an always-VLM baseline. AEGIS is open source and, to our knowledge, the first system to unify pre-flight assay-aware validation with runtime visual monitoring for an open-source liquid handler.

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 6 linked inside Pith

[1]

the robot ran

DRAFT - R X U Q D O 1 D P H AEGIS: Assay-Aware Protocol Validation and Runtime Monitoring for Open-Source Liquid Handling Robots † Priyanka V. Setty, ∗a,b Arvind Ramanathan, a Ian Foster, a,b and Rick Stevens a,b Self-driving laboratories increasingly rely on low-cost liquid handling robots such as the Open- trons OT-2, but these platforms ship without th...

Pith/arXiv arXiv 2026
[3]

is the robust fix. Fourth, the monitor-side latency of∼9.5 s can exceed the next protocol-level boundary in protocols that share a tip across consecutive destination wells; AEGIS still halts the run but would not prevent immediate-next-well contamination in those cases. Fifth, the live demonstration is on a single OT-2 across three planted-failure modes a...

2024
[4]

Skreta, N

11 M. Skreta, N. Yoshikawa, S. Arellano-Rubach, Z. Ji, L. B. Kris- tensen, K. Darvish, A. Aspuru-Guzik, F. Shkurti and A. Garg, arXiv preprint arXiv:2303.14100,

Pith/arXiv arXiv
[6]

13 J. O. Curtis, D. Ricketts, S. Mansoor, C. Cremer and P. Brisk, Proceedings of the ACM on Programming Languages (OOP- SLA), 2018, pp. 1–31. 14 B. Hsu, P. Setty , R. Butler, R. Lewis, C. Stone, R. Weinberg, T. Brettin, R. Stevens, I. Foster and A. Ramanathan,Digital Discovery,

2018
[8]

18 Y. Yin, J. Lei and W. Tao,Electronics, 2024,13,

2024
[11]

21 M. Ho, M. F. Ginting, I. R. Ward, A. Reinke, M. J. Kochender- fer, A.-a. Agha-Mohammadi and S. Omidshafiei,arXiv preprint arXiv:2602.16182,

arXiv
[12]

Hafner, J

22 D. Hafner, J. Pasukonis, J. Ba and T. Lillicrap, arXiv preprint arXiv:2301.04104,

Pith/arXiv arXiv
[14]

Bergmann, M

26 P. Bergmann, M. Fauser, D. Sattlegger and C. Steger, IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2019, pp. 9592–9600. 27 K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox and P. Gehler, IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2022, pp. 14318–14328. 28 Z. Zhang, C. Yue, H. Xuet al.,arXiv p...

Pith/arXiv arXiv 2019
[15]

29 S. Bai, W. Song, J. Chenet al.,arXiv preprint arXiv:2512.22983,

arXiv
[384]

Detection is L1+L2 verdict agreement on each protocol

12 | 1–14 + P V S O B M / B N F < Z F B S > < W P M > DRAFT Table3 Academy fleet-scaling sweep on the 16-protocol benchmark, sim- ulated OT-2 backends. Detection is L1+L2 verdict agreement on each protocol. Nrobots Wall (s) Speedup % ideal Detection 1 157.5 1.00×100% 15/16 2 86.0 1.83×92% 15/16 4 45.4 3.47×87% 15/16 8 29.2 5.39×67% 15/16 16 20.17.84×49% 1...

Pith/arXiv arXiv
[2023]

O’Donoghue, A

12 O. O’Donoghue, A. Shtedritski, J. Ginger, R. Abboud, A. Ghareeb, J. Booth and S. G. Rodriques,arXiv preprint arXiv:2310.10632,

Pith/arXiv arXiv
[2024]

Darvish, A

25 K. Darvish, A. Sohal, A. Mandal, H. Fakhruldeen, N. Radulov, Z. Zhou, S. Veeramani, J. Choi, S. Han, B. Zhang, J. Chae, A. Wright, Y. Wang, H. Darvish, Y. Zhao, G. Tom, H. Hao, M. Bogdanovic, G. Pizzuto, A. I. Cooper, A. Aspuru-Guzik, F. Shkurti and A. Garg,arXiv preprint arXiv:2601.13232,

arXiv
[2025]

8 Opentrons Labworks,Opentrons AI Protocol Designer and Pro- tocol Library,https://opentrons.com/ai, 2024, Accessed 2026-05-20. 9 A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White and P. Schwaller,Nature Machine Intelligence, 2024,6, 525–535. 10 K. Darvish, M. Skreta, Y. Zhao, N. Yoshikawa, S. Som, M. Bog- danovic, Y. Cao, H. Hao, H. Xu, A. Aspu...

2024
[2026]

hamiltoncompany.com/automated-liquid-handling/ platform/microlab-star-products/tadm, 2024, Accessed 2026-05-20

15 Hamilton Company ,Total Aspiration and Dispense Monitoring (TADM) Technology,https://www. hamiltoncompany.com/automated-liquid-handling/ platform/microlab-star-products/tadm, 2024, Accessed 2026-05-20. 16 Tecan Group,Process Monitored Pipetting (PMP) and Liquid- Level Detection on Tecan Fluent,https://www.tecan.com/ fluent-laboratory-automation, 2024, ...

2024
[2836]

19 Z. Zhou, S. Veeramani, F. Munguia-Galeano, H. Fakhruldeen and A. I. Cooper,Communications Chemistry, 2025,8,

2025

[1] [1]

the robot ran

DRAFT - R X U Q D O 1 D P H AEGIS: Assay-Aware Protocol Validation and Runtime Monitoring for Open-Source Liquid Handling Robots † Priyanka V. Setty, ∗a,b Arvind Ramanathan, a Ian Foster, a,b and Rick Stevens a,b Self-driving laboratories increasingly rely on low-cost liquid handling robots such as the Open- trons OT-2, but these platforms ship without th...

Pith/arXiv arXiv 2026

[2] [3]

is the robust fix. Fourth, the monitor-side latency of∼9.5 s can exceed the next protocol-level boundary in protocols that share a tip across consecutive destination wells; AEGIS still halts the run but would not prevent immediate-next-well contamination in those cases. Fifth, the live demonstration is on a single OT-2 across three planted-failure modes a...

2024

[3] [4]

Skreta, N

11 M. Skreta, N. Yoshikawa, S. Arellano-Rubach, Z. Ji, L. B. Kris- tensen, K. Darvish, A. Aspuru-Guzik, F. Shkurti and A. Garg, arXiv preprint arXiv:2303.14100,

Pith/arXiv arXiv

[4] [6]

13 J. O. Curtis, D. Ricketts, S. Mansoor, C. Cremer and P. Brisk, Proceedings of the ACM on Programming Languages (OOP- SLA), 2018, pp. 1–31. 14 B. Hsu, P. Setty , R. Butler, R. Lewis, C. Stone, R. Weinberg, T. Brettin, R. Stevens, I. Foster and A. Ramanathan,Digital Discovery,

2018

[5] [8]

18 Y. Yin, J. Lei and W. Tao,Electronics, 2024,13,

2024

[6] [11]

21 M. Ho, M. F. Ginting, I. R. Ward, A. Reinke, M. J. Kochender- fer, A.-a. Agha-Mohammadi and S. Omidshafiei,arXiv preprint arXiv:2602.16182,

arXiv

[7] [12]

Hafner, J

22 D. Hafner, J. Pasukonis, J. Ba and T. Lillicrap, arXiv preprint arXiv:2301.04104,

Pith/arXiv arXiv

[8] [14]

Bergmann, M

26 P. Bergmann, M. Fauser, D. Sattlegger and C. Steger, IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2019, pp. 9592–9600. 27 K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox and P. Gehler, IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2022, pp. 14318–14328. 28 Z. Zhang, C. Yue, H. Xuet al.,arXiv p...

Pith/arXiv arXiv 2019

[9] [15]

29 S. Bai, W. Song, J. Chenet al.,arXiv preprint arXiv:2512.22983,

arXiv

[10] [384]

Detection is L1+L2 verdict agreement on each protocol

12 | 1–14 + P V S O B M / B N F < Z F B S > < W P M > DRAFT Table3 Academy fleet-scaling sweep on the 16-protocol benchmark, sim- ulated OT-2 backends. Detection is L1+L2 verdict agreement on each protocol. Nrobots Wall (s) Speedup % ideal Detection 1 157.5 1.00×100% 15/16 2 86.0 1.83×92% 15/16 4 45.4 3.47×87% 15/16 8 29.2 5.39×67% 15/16 16 20.17.84×49% 1...

Pith/arXiv arXiv

[11] [2023]

O’Donoghue, A

12 O. O’Donoghue, A. Shtedritski, J. Ginger, R. Abboud, A. Ghareeb, J. Booth and S. G. Rodriques,arXiv preprint arXiv:2310.10632,

Pith/arXiv arXiv

[12] [2024]

Darvish, A

25 K. Darvish, A. Sohal, A. Mandal, H. Fakhruldeen, N. Radulov, Z. Zhou, S. Veeramani, J. Choi, S. Han, B. Zhang, J. Chae, A. Wright, Y. Wang, H. Darvish, Y. Zhao, G. Tom, H. Hao, M. Bogdanovic, G. Pizzuto, A. I. Cooper, A. Aspuru-Guzik, F. Shkurti and A. Garg,arXiv preprint arXiv:2601.13232,

arXiv

[13] [2025]

8 Opentrons Labworks,Opentrons AI Protocol Designer and Pro- tocol Library,https://opentrons.com/ai, 2024, Accessed 2026-05-20. 9 A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White and P. Schwaller,Nature Machine Intelligence, 2024,6, 525–535. 10 K. Darvish, M. Skreta, Y. Zhao, N. Yoshikawa, S. Som, M. Bog- danovic, Y. Cao, H. Hao, H. Xu, A. Aspu...

2024

[14] [2026]

hamiltoncompany.com/automated-liquid-handling/ platform/microlab-star-products/tadm, 2024, Accessed 2026-05-20

15 Hamilton Company ,Total Aspiration and Dispense Monitoring (TADM) Technology,https://www. hamiltoncompany.com/automated-liquid-handling/ platform/microlab-star-products/tadm, 2024, Accessed 2026-05-20. 16 Tecan Group,Process Monitored Pipetting (PMP) and Liquid- Level Detection on Tecan Fluent,https://www.tecan.com/ fluent-laboratory-automation, 2024, ...

2024

[15] [2836]

19 Z. Zhou, S. Veeramani, F. Munguia-Galeano, H. Fakhruldeen and A. I. Cooper,Communications Chemistry, 2025,8,

2025