Technical Summary: Task-Aware Scanning Parameter Configuration for Robotic Inspection

Problem Definition

Robotic laser profiling is a standard method for dimensional verification and surface inspection in manufacturing. However, the fidelity of measurements is often determined not by the robot's motion planning, but by the configuration of the sensor's operating parameters (e.g., sampling frequency, measurement range, exposure time, dynamic range, and illumination). Currently, these parameters are tuned manually via trial-and-error or via task-agnostic auto-gain functions that maximize signal strength without considering specific inspection intents (e.g., global outline vs. fine defect detection).

Incorrect configurations lead to irreversible data loss, such as saturation, clipping, or missing returns, which downstream algorithms cannot recover. Furthermore, inspection objectives are typically communicated via natural language instructions (e.g., "scan the top surface for fine details"), yet current robotic systems treat sensor configuration as a static setting rather than an adaptive decision variable conditioned on both the visual scene and the task intent. The paper identifies a gap in existing research: while active perception focuses on where to look, it rarely addresses how to sense based on task semantics and visual appearance.

Methodology: ScanHD

To address this, the authors propose ScanHD, a framework based on Hyperdimensional Computing (HDC) for instruction-conditioned sensing parameter recommendation. The system takes a pre-scan RGB observation and a natural-language inspection instruction to infer a discrete configuration for a laser profiler.

1. Dataset: Instruct-Obs2Param

The authors introduce a real-world multimodal dataset to benchmark this problem.

• Hardware: A UR3 manipulator equipped with a Keyence LJ-X8200 laser profiler and an Intel RealSense D435i camera.
• Content: 16 representative objects (electronics, tools, calibration blocks) across 16 objects, 3 lighting conditions, and 3 poses/rotations.
• Structure: The dataset links inspection intents (encoded as structured semantic slots), visual observations, and ground-truth discrete scanner parameters.
• Generation: A "Data Evolution Flywheel" is used to synthesize high-quality instruction-observation-parameter triplets, utilizing LLMs for instruction generation and expert-in-the-loop calibration to ensure physical validity and consistency.

2. The ScanHD Framework

ScanHD employs a representation-first approach using HDC to ensure robustness, interpretability, and low latency.

• Encoding:
  • Visual observations and natural-language instructions are embedded using a pre-trained Vision-Language Model (CLIP).
  • These embeddings are bound together to form an instruction-conditioned representation.
  • The representation is projected into a high-dimensional bipolar space (hypervectors) via a random projection encoder.
• Learning (Parameter-wise Associative Reasoning):
  • Instead of a monolithic end-to-end model, ScanHD uses parameter-specific associative memories. Each of the five scanning parameters has its own set of class hypervectors.
  • Training involves a single-pass adaptive learning phase followed by optional iterative refinement to sharpen decision boundaries for misclassified samples.
  • The bundling and binding operations of HDC allow for the compositional fusion of multimodal inputs while maintaining robustness to noise.
• Inference:
  • During deployment, the encoded query hypervector is compared against the learned class hypervectors using cosine similarity.
  • The parameter value with the highest similarity is selected via an argmax operation.
  • This process is lightweight, interpretable, and suitable for real-time robotic control loops.

Key Contributions

1. Problem Formulation: The paper reframes sensor configuration from a static engineering setting to a first-class, task-aware decision variable in embodied inspection, conditioned on both visual observation and natural language intent.
2. Instruct-Obs2Param Dataset: A real-world multimodal dataset linking inspection intents, visual observations, and discrete laser scanning parameters across diverse objects and conditions, addressing the lack of standardized benchmarks for this specific problem.
3. ScanHD Framework: A hyperdimensional computing framework that enables robust, interpretable, and parameter-wise reasoning. It binds multimodal inputs into symbolic representations and performs associative retrieval for low-latency inference.
4. Empirical Validation: Comprehensive experiments demonstrating that ScanHD outperforms rule-based heuristics, conventional multimodal models (ResNet, ViT, Early/Late Fusion), and Multimodal Large Language Models (MLLMs) in accuracy and robustness.

Experimental Results

Evaluated on the Instruct-Obs2Param dataset, ScanHD achieved the following:

• Accuracy: 92.7% average Exact Accuracy and 98.1% average Win@1 Accuracy across the five parameters.
• Comparison:
  • vs. Rule-based: Significantly outperformed rule-based heuristics (50.4% avg Exact Accuracy), particularly on parameters requiring visual grounding.
  • vs. Unimodal Models: Instruction-only models failed on appearance-sensitive parameters (e.g., exposure time), while observation-only models failed on intent-driven parameters (e.g., sampling frequency). ScanHD successfully fused both modalities.
  • vs. MLLMs: Multimodal LLMs (Qwen3-VL variants) performed poorly (<50% Exact Accuracy), struggling to map free-form reasoning to precise, hardware-constrained discrete choices without domain-specific inductive bias.
• Generalization: The model demonstrated strong cross-split generalization across position, rotation, and lighting changes, maintaining >91% Exact Accuracy in most split scenarios.
• Data Efficiency: The model achieved >90% Win@1 Accuracy with only 20% of the training data, indicating high data efficiency.
• Efficiency: ScanHD exhibited the lowest inference latency among all tested methods, making it suitable for real-time deployment, whereas MLLMs showed latency orders of magnitude higher.

Significance and Claims

The paper claims that ScanHD enables autonomous, instruction-conditioned sensing configuration, effectively eliminating the need for manual tuning and elevating sensor configuration from a static setting to an adaptive decision variable.

The authors emphasize that their approach addresses the "sensing-first" reality of industrial inspection, where errors introduced at the acquisition stage are irreversible. By using Hyperdimensional Computing, the system achieves a balance between robustness (handling appearance variations), interpretability (transparent decision pathways), and efficiency (low-latency inference), which are critical requirements for deployment in industrial robotic cells. The work validates that task-aware sensing adaptation is a distinct and solvable learning problem that current end-to-end deep learning and LLM approaches are not optimally suited to address without specific structural constraints.