Task-Specified Active Metrological Inspection with Measurement-Steered VLA Manipulation and Deterministic Evidence Gating

Zhiling Chen; Jingzhan Ge; Ruimin Chen; Matthew P. Castanier; David Gorsich; Farhad Imani

cs.ROSep 13, 2026

Task-Specified Active Metrological Inspection with Measurement-Steered VLA Manipulation and Deterministic Evidence Gating

Authors: Zhiling Chen, Jingzhan Ge, Ruimin Chen, Matthew P. Castanier, David Gorsich, Farhad Imani

Organizations: University of Connecticut, Storrs, CT, USA · U.S. Army DEVCOM Ground Vehicle Systems Center (GVSC), Warren, MI, USA

Abstract

High-mix low-volume (HMLV) manufacturing requires inspection systems to adapt to changing parts, specifications, and work orders without repeated task-specific programming. Existing inspection automation typically assumes predefined sensing sequences, while general purpose robot agents optimize task completion rather than the completeness and validity of metrological evidence. We formulate task-specified active metrological inspection and propose From Requirements to Admissible Metrological Evidence (FRAME), a hierarchical dual-arm framework that converts an inspection instruction and structured specification into traceable conformance evidence. FRAME coordinates learned manipulation with calibrated laser profilometry: a task manager grounds and schedules requirements, active surface correspondence verifies physical-to-specification localization, and evidence memory tracks measurement provenance, admissibility, and coverage. Learned components may propose inspection targets and physical access actions, but deterministic datum-grounded measurement, admissibility checks, coverage auditing, and conformance evaluation prevent incomplete or unverified evidence from authorizing PASS. A series of physical experiments shows that FRAME achieves higher end-to-end inspection reliability, fewer false accepts, and shorter task completion time.

Figures & tables

Fig. 1: Motivation and overview of task-specified active inspection for HMLV manufacturing. High part and specification variability makes conventional inspection preparation costly, while existing inspection technologies trade off accuracy, coverage, and per-part setup effort. A stated inspection task may require features distributed across visible, partially occluded, and hidden surfaces, motivating both task-specified measurement and active reorientation. Our system takes a natural-language instruction and structured specification, grounds the required features to the current part pose, alternates look , scan , and reorient to acquire verified evidence, produces a deterministic conformance verdict, and invokes sort for an authorized binary outcome.

Fig. 2: Overview of our framework. The specification-constrained task manager coordinates a manipulation expert and a measurement expert through shared Surface Correspondence and Evidence Memory. The manipulation expert exposes three atomic skills: pre-calibrated look , VLA-based reorient , and VLA-based sort . Each reorientation is independently verified by look and the correspondence gate before scanning. The measurement expert converts calibrated profiles into admissible geometric evidence, while failed acquisitions trigger reacquire . Complete coverage enables deterministic conformance evaluation: a supported pass or no-pass verdict triggers sort , whereas exhausted execution or reacquisition budgets produce unresolved and manual review.

Fig. 3: Specification-driven construction of the surface-correspondence table. Step 1: Minimum LOOK-set compilation. Given the part specification and its geometry-family LOOK library, the compiler selects the smallest subset that separates all non-equivalent stable configurations; in the cube example, one upper LOOK is ambiguous, whereas upper and left LOOKs identify one configuration. Step 2: Per-view region evidence acquisition. The robot executes the selected calibrated LOOKs, SAM3 [ 38 ] crops the workpiece, and the VLM scores the candidate specification regions. Step 3: Correspondence construction. The per-view scores rank the legal configurations. If the best candidate passes the score, top-two-margin, and anchor/topology gates, its physical-to-specification assignment is released as $\widehat{\mu}_{t}$ ; otherwise, no table is produced.

Fig. 4: Measurement-range configuration of the Keyence LJ-X8200 laser profiler. The measurement window can be shifted along the $Z$ -axis by adjusting its center position and reduced from the full 68 mm range to 34, 17, 8.5, or 4.25 mm. Reducing the measurement range increases the maximum sampling frequency $f_{\max}$ from 500 Hz to 8 kHz.

Fig. 5: Examples of specification-driven vertical measurement-window selection for different objects and inspection requests. The instruction is grounded to the required surface, feature depth or height, and tolerance before the expert selects the measurement $Z$ range and center position. The displayed sampling frequency is the maximum frequency $f_{\max}(R_{f})$ supported by the selected $Z$ range, rather than the commanded scan frequency. (a) The cube bolt pattern uses a narrow range centered near the raised surface. (b) The shallow groove on the bracket uses the same range size but a shifted center because its base surface is at a different height. (c) The counterbore on the high tread again uses a narrow, surface-centered range. (d) Step-height inspection must contain both height levels, requiring a wider range, an intermediate center, and consequently a lower maximum sampling frequency. The acquired profiles confirm that the task-required geometry lies within the configured window.

Fig. 6: Measurement Expert pipeline from a verified inspection request to admissible evidence. (a) The grounded specification defines the feature region, lateral scan path, and vertical measurement window. (b) Calibrated laser profiles are registered into a height map. (c) Host-surface tilt is normalized before geometric candidates are extracted. (d) The datum reference frame is constructed and candidates are deterministically associated with the requested specification features. (e) A requirement-specific operator computes the reported geometric quantity. (f) Deterministic admissibility predicates either admit the measurement to Evidence Memory or trigger targeted reacquisition; an inadmissible record cannot support conformance.

Fig. 7: Our real-world experiment setup. One XArm 6 carries a Keyence LJ-X8200 laser line profiler, while the other is equipped with a G2 gripper and wrist camera for part manipulation and local observation. A fixed top-view camera provides global workspace observations. Parts are measured in the inspection zone and subsequently sorted into the pass or no-pass zone.

Task	Category	Industrial need	Required evidence
$T_{1}$	Specified feature	Verify a named feature after machining or printing	Measure required characteristic of the requested feature
$T_{2}$	Anomaly detection	Protect mating, sealing, bearing, or datum surfaces	Measure damage and form error on the requested surface
$T_{3}$	Relation or GD&T	Ensure alignment, interchangeability, and compatibility	Evaluate relations among features, surfaces, and datums
$T_{4}$	Composite inspection	Inspect a declared feature group or interface	Aggregate all requirements in the declared group
$T_{5}$	Functional goal	Verify an intent-level production goal	Infer the required scope and collect complete evidence

TABLE I: Inspection task taxonomy and its industrial motivation. Manipulation difficulty is evaluated independently through reorientation cost and regret.

Fig. 8: Designed 3D-printed benchmark objects across increasing geometric variation. Four object families, bracket, cylinder, block, and stepped block, are organized into three levels. Level 0 provides the nominal geometries used for manipulation policy training, Level 1 introduces feature-level variation while largely preserving the overall shape, and Level 2 introduces larger morphology changes, with two variants per family.

Fig. 9: End-to-end reliability and task completion cost. (a) Paired end-to-end success across L1, L2, and CNC parts, where a pair is counted as successful only when both the matched PASS-spec and FAIL-spec episodes are completed correctly. (b) Mean task-to-completion time, measured from task-specific preparation to final report and sorting or unresolved termination. (c) Terminal outcomes on PASS-spec episodes, reported as correct PASS, false reject, and unresolved. (d) Terminal outcomes conditional on FAIL-spec episodes, reported as correct FAIL, conditional false accept, and unresolved. An episode is unresolved when no valid binary verdict is produced within the common execution and recovery budget and is therefore transferred to manual review.

Fig. 10: Metrology validation across representative measurement operators. (a) Signed bias, mean absolute error (MAE), and 95th-percentile absolute error (P95) relative to reference measurements for opening diameter, depth/surface offset, groove footprint dimensions, projected location/coaxiality, and form/relative orientation. The dashed vertical line denotes zero signed bias. (b) Repeatability standard deviation over repeated measurements for the same operator groups, quantifying scan-to-scan measurement consistency.

Method	Cand. acc. (%) $\uparrow$	Wrong rel. / trial (%) $\downarrow$	Abstain (%)	Rel. acc. (%) $\uparrow$
L1
VLM w/o FRAME	54.8	45.2	-	-
VLM w/ FRAME	92.9	2.4	9.5	97.4
L2
VLM w/o FRAME	52.4	47.6	-	-
VLM w/ FRAME	82.1	3.6	19.0	95.6

TABLE II: LOOK-based surface-correspondence performance. Both methods receive the same fixed LOOK observations; FRAME additionally resolves per-view VLM scores over legal configurations and applies the confidence–consistency gate before releasing a correspondence.

Fig. 11: Scope risk evaluation with and without deterministic scope expansion (DSE). Correct grounding rates are reported at three levels of the specification hierarchy: (a) required faces, (b) corresponding features, and (c) leaf requirements, across task types T1–T5.

Method	Correct-face acceptance $\uparrow$	Wrong-face admission $\downarrow$
w/o correspondence gate	15/15	25/25
w/ correspondence gate	14/15	2/25

TABLE III: Post-reorientation state admission on matched L2 executions with and without the correspondence gate. Rejected and abstained states are counted as not admitted.

Fig. 12: Representative metrology failures and recovery. Three cases illustrate Z-range clipping, incomplete lateral coverage, and incorrect host-plane selection. Requirement-specific admissibility checks reject the invalid acquisition and trigger targeted recovery before the measurement is admitted as evidence.

Fig. 13: Case study of cross-face groove inspection. Given the instruction and structured specification, FRAME expands the required scope from target faces to groove features and leaf requirements, including width, length, depth, floor parallelism, and the cross-face size relation. The system verifies the first surface before scanning, admits only valid measurements into Evidence Memory, then reorients the part, re-verifies the exposed surface, and performs the second scan. After complete evidence coverage is achieved, the deterministic conformance evaluator identifies a groove-size mismatch and issues a NO-PASS verdict before sorting.

Robotic Inspection

Robotic inspection uses autonomous or remotely operated robots to examine infrastructure, equipment, and environments, detect defects, and document condition.

See all 50 papers

Task-Aware Scanning Parameter Configuration for Robotic Inspection Using Vision Language Embeddings and Hyperdimensional Computing
May 5, 2026Zhiling Chen, David Gorsich, Matthew P. Castanier +3Robotic Inspection Robotics
aSchool of Mechanical, Aerospace, and Manufacturing Engineering, University of Connecticut, Storrs, Connecticut, USA · bUS Army DEVCOM Ground Vehicle Systems Center, Warren, Michigan, USA
Admittance-Based Surface Alignment for Human-in-the-Loop Robotic Visual Inspection
Jun 17, 2026Antara Banerjee, Colin Acton, Xu ChenRobotic Control Robotic Inspection
Department of Electrical and Computer Engineering, University of Washington, Seattle · Department of Mechanical Engineering, University of Washington, Seattle
Illumination-Aware Active Perception for Spacecraft Inspection
Oct 15, 2026Sagarika Rao Valluri, Benjamin RiviereRobotic Inspection Active Visual Perception
Mechanical and Aerospace Engineering, New York University · Mechanical and Aerospace Engineering and Computer Science, New York University

May 5, 2026J/K move · Enter open · S save

Task-Aware Scanning Parameter Configuration for Robotic Inspection Using Vision Language Embeddings and Hyperdimensional Computing

Zhiling Chen, David Gorsich, Matthew P. Castanier, Yang Zhang+2

aSchool of Mechanical, Aerospace, and Manufacturing Engineering, University of Connecticut, Storrs, Connecticut, USA · bUS Army DEVCOM Ground Vehicle Systems Center, Warren, Michigan, USA

Robotic Inspection Robotics Hyperdimensional Computing VLMs for Robotics Robotic Perception

Open paper AskPDF arXiv

Robotic laser profiling is widely used for dimensional verification and surface inspection, yet measurement fidelity is often dominated by sensor configuration rather than robot motion. Industrial profilers expose multiple coupled parameters, including sampling frequency, measurement range, exposure time, receiver dynamic range, and illumination, that are still tuned by trial-and-error; mismatches can cause saturation, clipping, or missing returns that cannot be recovered downstream. We formulate instruction-conditioned sensing parameter recommendation; given a pre-scan RGB observation and a natural-language inspection instruction, infer a discrete configuration over key parameters of a robot-mounted profiler. To benchmark this problem, we develop Instruct-Obs2Param, a real-world multimodal dataset linking inspection intents and multi-view pose and illumination variation across 16 objects to canonical parameter regimes. We then propose ScanHD, a hyperdimensional computing framework that binds instruction and observation into a task-aware code and performs parameter-wise associative reasoning with compact memories, matching discrete scanner regimes while yielding stable, interpretable, low-latency decisions. On Instruct-Obs2Param, ScanHD achieves 92.7% average exact accuracy and 98.1% average Win@1 accuracy across the five parameters, with strong cross-split generalization and low-latency inference suitable for deployment, outperforming rule-based heuristics, conventional multimodal models, and multimodal large language models. This work enables autonomous, instruction-conditioned sensing configuration from task intent and scene context, eliminating manual tuning and elevating sensor configuration from a static setting to an adaptive decision variable.

Abstract

Robotic Inspection

Explore similar work

Task-Aware Scanning Parameter Configuration for Robotic Inspection Using Vision Language Embeddings and Hyperdimensional Computing

Admittance-Based Surface Alignment for Human-in-the-Loop Robotic Visual Inspection

Illumination-Aware Active Perception for Spacecraft Inspection