TL;DR
This paper presents an automated pipeline to bridge the gap between human-readable disassembly manuals and the structured data required for robotic automation. By comparing a rule-based NLP parser (spaCy) against a generative AI approach (Gemini 2.5 Flash), the researchers demonstrate that while LLMs excel at understanding complex actions and tools, traditional rule-based methods remain superior for precise data extraction and processing speed.
What Problem Does This Paper Address?
As the global economy shifts toward "circularity"—emphasizing the repair, refurbishment, and recycling of products (often called "Re-X" strategies)—disassembly becomes a critical bottleneck. Large companies can invest in custom automation, but Small and Medium-sized Enterprises (SMEs) struggle with the high cost of planning disassembly for a diverse range of complex products.
The primary hurdle is data. Disassembly instructions usually live in unstructured PDF manuals designed for humans. To automate these tasks, robots need structured "Disassembly Sequence Planning" (DSP) models. Currently, converting a PDF manual into a machine-readable format requires extensive manual effort. This paper explores how to automate that "knowledge acquisition" phase using modern AI.
Methods
The authors developed a multi-stage pipeline to transform raw PDFs into structured tables containing six categories: Action, Object, Tool, Count, Condition, and Position.
The pipeline consists of:
- Pre-processing: Converting PDFs to Markdown using the Docling library and using Gemini 2.5 Flash to filter out "noise" like legal disclaimers or warranty info.
- The NLP Parser (spaCy): A rule-based approach using part-of-speech (POS) tagging and dependency trees. It relies on strict linguistic patterns (e.g., "use [tool] to [verb]") to identify instruction components.
- The GenAI Extractor (Gemini 2.5 Flash): A generative approach using Chain-of-Thought (CoT) prompting and few-shot examples to "reason" through the text and extract the same six categories.
Datasets
The researchers created a new benchmark dataset for this task:
- Source: 22 real-world disassembly manuals covering consumer electronics (monitors), industrial machinery (bearings), and household appliances.
- Ground Truth: 741 instructions manually annotated by humans across the six defined categories.
- Complexity: The dataset was split into "simple" and "complex" documents based on the density of tool, condition, and position requirements.
Key Insights and Ideas
- The "Context vs. Precision" Trade-off: The GenAI extractor was significantly better at identifying Actions and Tools. This is because LLMs handle complex phrasal verbs (like "snap off") and technical jargon much better than a rigid rule-set.
- Rules Still Rule for Structure: The rule-based NLP parser outperformed the LLM in the Object and Count categories. Why? The NLP parser was programmed to split conjoined objects (e.g., "remove the screw and the washer") into separate instructions deterministically, whereas the LLM tended to keep them as a single natural language string.
- Speed is a Factor: The NLP parser is nearly 27 times faster than the GenAI approach. For massive industrial databases, this latency difference is non-trivial.
- The "Exact Match" Challenge: Even with state-of-the-art LLMs, the "Exact Match" (EM) rate for the NLP parser was only ~50%, and the GenAI was even lower (~37%). This suggests that while AI can find the information, perfectly formatting it for a robot is still a very hard problem.
Key Results
The study found that both frameworks achieved remarkably similar F1 scores (~0.89), but their strengths were complementary.
- Precision: The GenAI Extractor had higher precision (0.93 vs 0.91), meaning its predictions were less likely to be "hallucinations" or errors.
- Consistency: The NLP Parser achieved a much higher Exact Match (EM) score (0.50 vs 0.36), indicating it was more consistent at filling out all six categories correctly for a given instruction.
- Efficiency: The NLP parser processed nearly 30 instructions per second, while the GenAI managed only 1.1 instructions per second.
Limitations and Future Work
The authors acknowledge that text alone isn't enough. Real disassembly manuals are full of diagrams, technical drawings, and photos that provide spatial context text cannot capture.
Future directions include: * Multimodal Extraction: Integrating computer vision to process illustrations and CAD data alongside text. * Semantic Mapping: Moving beyond just extracting text to matching that text to a "canonical" library of robotic actions (e.g., mapping "unscrew," "rotate," and "loosen" to a specific robot primitive). * Hybrid Models: Combining the linguistic flexibility of LLMs with the deterministic precision of rule-based parsers to improve the Exact Match rate.
Summary updated