LabBuilder: Protocol-Driven 3D Lab Layout Generation from Text Descriptions (ICML 2026)

Automated laboratories promise to accelerate scientific research, but designing virtual lab environments remains a core bottleneck for large-scale deployment. Existing 3D scene generation methods target residential scenes and lack three critical capabilities for research labs: semantic understanding of experimental protocols (instrument and reagent placement must follow specific reaction workflows), chemical safety constraints (flammable reagent separation from heat sources, incompatible chemical isolation), and robot reachability (autonomous robots must navigate to each operation station in protocol order without collisions). Moreover, current methods optimize for visual plausibility during generation, deferring protocol feasibility and navigation validation to post-hoc evaluation, often yielding layouts with safety or functional defects.

To address these gaps, researchers from Shanghai AI Laboratory, Wuhan University, Beihang University, Peking University, Tsinghua University, Shanghai Jiao Tong University, and Chinese University of Hong Kong jointly proposed LabBuilder , accepted at ICML 2026. Given a free-text experiment description, LabBuilder produces an end-to-end 3D lab layout that is geometrically compliant, chemically safe, and robot-navigable, turning protocol-driven lab design into a fully automated process.

Core Contributions

Meta-Lab Dataset : An extensible dataset comprising an Asset Knowledge Base (176 lab entity classes with geometry, semantic category, and chemical hazard annotations including flammability ratings and incompatibility lists) and a Chemical Knowledge Base (six reaction types extracted from real experimental data: substitution, cyclization, redox, functional-group transformation, condensation, alkylation).

LabGen Hierarchical Layout Generation : A multi-stage framework that generates protocol-driven layouts under geometric, chemical, and navigation constraints, outperforming existing methods.

LabTouchstone Evaluation Benchmark : A four-dimensional assessment suite — geometric compliance (Obj↑, OB↓, CN↓), feasibility success rate (FSR: Asset↑, Nav↑), chemical safety (Flam↑, Store↑, Incomp↑, Glass↑), and semantic plausibility (Real↑, Lay↑, Comp↑ via LLM scoring) — plus protocol-derived point-goal navigation evaluation for standardized cross-method comparison.

System Architecture

LabBuilder consists of three modules:

LabForge: Asset & Protocol Compiler

LabForge builds the knowledge bases and synthesizes structured protocols. Using an LLM core, it compiles free-text experiment descriptions into structured protocols containing operation-step sequences, required asset lists, and safety constraint sets. In 30 annotated experiments, each protocol averaged 5.27 reagent types, 9.87 instrument types, 9.00 operation steps, and 4.30 navigation actions, reflecting the complexity of lab scenes in asset variety and procedural flow.

LabGen: Hierarchical Layout Generation & Optimization

LabGen employs a phased generation and iterative optimization pipeline:

Layout Initialization : Two-level hierarchy — Room-Level Zoning identifies protocol-required functional zones, assigns 6-DoF poses, and allocates space; Desktop-Level Organization fills protocol-relevant assets onto work surfaces under boundary and non-overlap constraints to produce candidate initial layouts.

Joint Geometric & Chemical Optimization : Optimizes a multi-objective reward function F measuring geometric validity (boundary and collision constraints) and chemical safety (hazard-annotation constraint satisfaction). Two complementary mechanisms alternate at room and desktop levels until convergence or max iterations: FastRepair quickly resolves geometric boundary conflicts and collisions; LLMAdjust invokes an LLM for fine-grained safety and workflow constraint adjustment.

Navigation-Aware Optimization : Projects the 3D scene to a 2D occupancy grid with obstacle inflation, validates path feasibility for each protocol step using A* search, and detects three failure modes — invalid endpoints, boundary violations, unreachable paths. Scenes with navigation failures undergo iterative optimizer adjustment until all navigation constraints are satisfied or iteration limit reached.

LabTouchstone: Four-Dimensional Evaluation

Quantifies layout quality across geometric compliance, feasibility success rate (asset availability and protocol-level navigation feasibility), chemical safety (flammable isolation, storage compliance, incompatibility separation, glassware handling), and semantic plausibility (realism, layout quality, completeness via LLM scoring). Supports protocol-derived point-goal navigation evaluation for standardized quantitative comparison.

Experimental Results

Quantitative Comparison (30 Chemical Protocols)

LabBuilder outperforms all baseline scene-generation methods across every metric. Notably, it achieves a navigation success rate (FSR-Nav) of 0.966 and is the only method supporting protocol-level navigation validation. Detailed metrics show superior geometric compliance, chemical safety scores, and semantic quality.

Navigation Performance

Testing iPlanner and NavDP planners in LabBuilder-generated scenes shows both planners' success rates and path quality improve markedly after navigation-aware optimization, confirming practical usability for autonomous robots.

Ablation Study

Removing protocol guidance causes the largest drop: FSR-Asset falls from 0.833 to 0.300, completeness score from 9.10 to 3.14, confirming correct asset sets and constraints are foundational.

Removing geometric & chemical optimization degrades chemical safety sharply: flammable isolation (Flam) drops from 0.725 to 0.182, incompatibility separation (Incomp) from 0.716 to 0.367.

Removing navigation-aware optimization mainly reduces protocol-level navigation success (FSR-Nav from 0.966 to 0.870) with minor impact on geometric/chemical metrics, confirming its focused role in reachability assurance.

Limitations & Future Work

LabBuilder currently does not validate equipment-specification-level requirements (heating ranges, pipette capacities, fume-hood airflow), does not cover robot physical manipulation feasibility, lacks multi-protocol lab design support, and sim-to-real transfer remains open. Future directions include operation-spec verification, broader regulatory constraints, and multi-protocol laboratory design.

Conclusion

LabBuilder unifies experimental protocol semantic understanding, chemical safety modeling, and robot navigation reachability into a single 3D lab layout generation framework, shifting safety and reachability from post-hoc verification to built-in generative constraints. It establishes an open, standardized evaluation benchmark. Experiments demonstrate that explicitly incorporating protocol constraints significantly improves layout quality in safety, functionality, and robot usability for research-lab scene generation.

Code repository: https://github.com/che-0212/LabBuilder

Project page: https://che-0212.github.io/LabBuilder-site/