Making Self-Driving Laboratories Auditable:A Unified Framework for Hybrid Actions,Measurements, and Safe Execution

Zihan Zhang; Zeyu Wu; Xin Zhang; Mingyu Luo; Haohui Que; Tong Zhu

cs.ROOct 27, 2026

Making Self-Driving Laboratories Auditable:A Unified Framework for Hybrid Actions,Measurements, and Safe Execution

Authors: Zihan Zhang, Zeyu Wu, Xin Zhang, Mingyu Luo, Haohui Que, Tong Zhu

Organizations: SII–ECNU Joint Autonomous Chemistry Laboratory, Shanghai Innovation Institute · Nankai University · Shanghai Lixun Intelligent Technology Co.,Ltd · Shanghai Jiao Tong University ·

Abstract

Self-driving laboratories (SDLs) are laboratory systems that couple experimental planning, automated execution, and measurement feedback in a closed loop. Integrating flexible robot control with the typed application programming interfaces (APIs) and continuous measurements exposed by laboratory instruments remains an execution-interface challenge. To address this challenge, we propose a traceable framework comprising four components: LAM (Lab Action Model), OpenSDL, Safe-SDL, and L3-Bench. LAM is a proposed measurement-conditioned policy architecture that extends vision-language-action paradigms to generate both continuous robot trajectories and typed instrument API calls. OpenSDL is an execution substrate with an initial control implementation, designed to map role-level operations to heterogeneous devices, coordinate execution, and record timestamped causal events. The proposed Safe-SDL runtime gate combines pre-dispatch schema and resource checks with signal temporal logic monitoring during execution, including extended waiting periods. L3-Bench (Level-3 Laboratory Benchmark) specifies an episode-level evaluation protocol covering physical outcomes, timing, reproducibility, safety violations, and fault recovery. Together, these components provide an end-to-end design for linking hybrid action generation, monitored device execution, measurement feedback, and experimental evaluation, addressing coordination and traceability challenges in SDL workflows.

Figures & tables

Figure 1: LAM architecture for measurement-conditioned hybrid laboratory control. Task language, RGB/RGB-D observations, robot proprioception $q_{t}$ , and time-stamped instrument telemetry $M_{t}$ enter a multimodal context builder comprising a vision-language-model (VLM) backbone and a telemetry encoder. Typed measurement tokens retain device roles, units, sampling times, and data-quality information within the shared context. A continuous action head generates robot-trajectory chunks, a discrete function-calling head generates typed instrument application programming interface (API) calls, and an optional dynamics head predicts the next measurement state $M_{t+1}$ . The execution layer connects the arm controller and the signal temporal logic (STL) safety monitor with event logging to the OpenSDL bus. Parallel dispatch coordinates compatible actions, while updated observations and confirmed execution events supply feedback for subsequent decisions.

Field group	Examples	Audit invariant
Identity	episode ID, event ID, parent IDs	Every action and return is traceable to one episode and its causal predecessors.
Observation	RGB/RGB-D, $q_{t}$ , $M_{t}$ , quality flags	Sensors carry units, calibration version, sample time, and stale/missing status.
Action	role, op, typed parameters, idempotency key	The submitted action is identical to the schema-validated action used for evaluation.
Execution	provider, protocol version, ack, completion, raw bytes	Device-level failures are separated from model errors and safety blocks.
Safety	formula version, robustness, gate decision, operator override	A blocked or approved action has an explicit, replayable reason.
Environment	asset commit, configuration hash, seed, clock offset	A replay can reconstruct the software, asset, and timing context.

Table 1: Minimum fields required for an auditable SDL-Traj event.

Figure 2: Overview of the L3-Bench evaluation protocol. Left: adaptive dilution, measurement-conditioned titration, and fault injection with recovery define three scenario families. Two baselines—a vision-only single-arm policy and LAM without measurement tokens—are compared with the full LAM $\times$ OpenSDL framework under nominal, measurement-shift, device-transfer, and fault-injection conditions. Center: initialization comprises profiling, matching, activation, and configuration, followed by on-the-fly schema, safety, and replay-readiness checks of manifest and ledger completeness. The safety-gate ablation is restricted to simulation or risk-assessed hardware-in-the-loop (HIL) evaluation. Right: liquid separation/heating and delay/high pressure illustrate task and condition metadata; the evaluated scenario families are defined on the left. Complete episodes are scored, with batch-level clustering for bootstrap confidence intervals. Primary endpoints are normalized physical yield $Y$ , temporal safety violation rate SVR, and operational time cost $C$ ; diagnostic measures include mass-transfer error $\Delta m$ , temperature-control error ITAE, makespan, and recovery. Photographs illustrate the laboratory scenarios; check marks and miniature plots indicate protocol requirements and metric types.

Evaluation setting	Controlled intervention	Primary question	Required report
Nominal	Same devices, protocol, and measurement range	Can the full loop execute the task?	$Y$ , $C$ , completion rate
Measurement shift	Noise, latency, or target-range change	Do measurement tokens improve adaptation ( RQ1 )?	$\Delta m$ , $CV$ , calibration error
Fault injection	Offset, slip, timeout, over-temperature	Does gating contain and recover from faults ( RQ3 )?	SVR, interception, recovery time
Device transfer	New provider with the same role descriptor	Are action semantics portable ( RQ4 )?	Schema failures, transfer gap

Table 2: Pre-specified evaluation matrix for L3-Bench.

Component	Status and evidence
OpenSDL 1.0 mother/edge/transport	Control engine, adapters, transports, and SQLite logging implemented; end-to-end replay validation pending.
Role/DeviceProvider and SDL-Traj	Proposed research contracts; full schema and compatibility validation pending.
LAM hybrid action heads	Proposed architecture and training objectives; trained checkpoints and ablations pending.
Clean-3D assets and digital twin	Asset specification; released manifests and validated scene inventory required.
L3-Bench	Metrics, scenarios, and evaluation protocol defined; full results pending.
Hardware transfer and safety statistics	Hardware-transfer experiments and safety statistics remain pending.

Table 3: Implementation status and evidence planned for the arXiv release.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Label	Operational definition
MODEL	LAM proposes an incorrect role, trajectory, or parameter despite valid inputs.
SCHEMA	Type, unit, range, or required-field validation rejects the proposal.
RESOURCE	A device, workspace, or barrier dependency cannot be acquired.
DEVICE	Provider returns a hardware, calibration, or communication failure.
SAFETY	STL, heartbeat, or data-quality gate blocks or aborts execution.
REPLAY	The episode cannot be reconstructed from its manifest and event ledger.

Appendix

Table 4: Failure labels used for episode-level analysis.

Laboratory Automation

Laboratory automation uses robotics, software, and AI to plan, execute, monitor, and analyze experiments with less manual intervention.

See all 62 papers

La Agente Óptima: Towards Agentic Self-Driving Laboratories
Sep 3, 2026Marcel Müller, Jiaru Bai, Willi Gottstein +14Bayesian Optimization Large Language Model-Guided Optimization
Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON2026 M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada +11
Self-Building Labs
Oct 16, 2026Jiaru Bai, Yuchi Zhao, Han Hao +4Laboratory Automation Autonomous Experimentation
Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada +6
PUDA: An AI-Native Hardware Harness for Self-Driving Laboratories
Jul 29, 2026Zekun Ren, Hongzhao Tan, Jiaen Yee +1Laboratory Automation Tool-Using Agents
Berkeley Education Alliance for Research in Singapore (BEARS) · School of Materials Science and Engineering, Nanyang Technological University (NTU)

Sep 3, 2026J/K move · Enter open · S save

La Agente Óptima: Towards Agentic Self-Driving Laboratories

Marcel Müller, Jiaru Bai, Willi Gottstein, Abhijoy Mandal+13

Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON2026 M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada+11

Bayesian Optimization Large Language Model-Guided Optimization LLM Agent Orchestration Laboratory Automation Autonomous Experimentation

Open paper AskPDF arXiv

Self-driving laboratories (SDLs) combine automated experimentation with adaptive decision-making to accelerate scientific discovery. Their operation nevertheless often depends on human specialists who translate scientific objectives into executable closed-loop campaigns. Specialists adjust them as data and operating conditions change. Here, we present La Agente Óptima, an agentic framework that constructs and supervises Bayesian optimization campaigns across computational and experimental systems while maintaining a persistent optimization state. By separating large language model (LLM) reasoning from executed campaigns, Óptima runs repetitive optimization loops consistently, returns control to the agent only when progress requires interpretation or campaign revision, and keeps every decision auditable. We evaluate Óptima across ablation studies, five digital discovery tasks, and two physical platforms. Throughout, Óptima maintained executable campaigns as both the scientific problem and execution environment evolved. In a closed-loop contact angle optimization campaign, Óptima identified and corrected a mid-run measurement failure, bringing the contact angle from 71.4 to 67.8 degrees, just above the 64-66 degree range. From this result, Óptima correctly inferred that the target was likely unattainable with the available reagents and recommended changing the formulation. In a five-day multi-objective flow-chemistry campaign, Óptima increased the yield from 30% to 59% over 23 experiments. Despite substantial inference costs, it cost less and used substantially less starting material than a human-directed campaign, while selecting a more mass-efficient operating point. These results show that LLM-based agents can make rigorous, long-running optimization campaigns accessible to domain scientists without specialist setup, expanding the scope of SDLs.

Abstract

Laboratory Automation

Explore similar work

La Agente Óptima: Towards Agentic Self-Driving Laboratories

Self-Building Labs

PUDA: An AI-Native Hardware Harness for Self-Driving Laboratories