Building a Reproducible Literature Search for Robotics and Engineering Research

Engineering and robotics researchers often treat the literature review as a preliminary chore: skim the obvious conference proceedings, collect a few dozen PDFs, and move on to the real work of building and testing systems. Yet reviewers increasingly expect related-work sections to be complete, current and fair, and fields such as robot learning, perception and human-robot interaction now publish so quickly that informal searching misses important work. This article describes a reproducible literature search workflow tailored to engineering research, covering where to search, how to document decisions, how to handle preprints and code, and how evidence-linked research assistants such as Yila can help keep related-work claims accurate.

Why reproducibility matters in a literature search

A reproducible search is one that another researcher could repeat and obtain broadly the same set of sources. It matters for three reasons.

First, it protects you during peer review. When a reviewer asks why a competing method was not compared, you can show that it fell outside your documented scope or appeared after your search date, rather than admitting it was missed.

Second, it saves time on revisions and follow-up papers. A documented search can be re-run months later to pick up new work, instead of starting again.

Third, it improves the quality of the related-work section itself. Explicit inclusion criteria force you to decide what counts as relevant, which sharpens the framing of your contribution.

Step one: define the scope precisely

Start with a one-sentence statement of the problem your work addresses, then list the dimensions along which related work can vary. For a paper on learning-based grasping of deformable objects, those dimensions might include:

• Object type: rigid, articulated, deformable, granular.
• Sensing: vision only, tactile, force-torque, multimodal.
• Learning approach: supervised, reinforcement, imitation, self-supervised.
• Evaluation: simulation only, real hardware, both.
• Time window: for example, the last eight years, with seminal earlier work added by hand.

Write down which combinations are in scope. This becomes your inclusion criteria.

Step two: choose sources suited to engineering

Engineering literature is spread across venues that general academic search engines index unevenly.

• Conference proceedings. In robotics, ICRA, IROS, RSS, CoRL and HRI are central; in machine learning, NeurIPS, ICML and ICLR. Many are indexed by IEEE Xplore, the ACM Digital Library or the conferences' own proceedings sites.
• Journals. IEEE Transactions on Robotics, the International Journal of Robotics Research, Robotics and Autonomous Systems, Science Robotics and others.
• Preprint servers. arXiv categories such as cs.RO, cs.LG and cs.CV carry much of the newest work, often months before formal publication.
• Open bibliographic indexes. OpenAlex and Semantic Scholar provide broad coverage, citation links and APIs useful for systematic searching and citation chaining.

Using at least one discipline-specific database, one open index and arXiv covers most of what matters.

Step three: build and log your search strings

Translate your scope into search strings with synonyms and Boolean operators, for example:

("deformable object" OR cloth OR cable OR rope OR garment) AND (grasp* OR manipulat*) AND (learn* OR "reinforcement learning" OR "imitation learning")

Keep a log with the database, date, exact string, filters and number of results. Adapt the syntax for each database and record the adaptations. This log is the core of reproducibility.

Step four: screen with explicit criteria

Screen titles and abstracts first, then full texts. For each excluded paper at the full-text stage, record a short reason such as "rigid objects only", "no learning component" or "simulation only, out of scope". A simple spreadsheet works well. For larger searches, dedicated screening tools can help, and two people screening independently for a sample of papers improves reliability.

Step five: handle preprints carefully

Preprints are essential in fast-moving fields but carry risks. Results may change before publication, and some preprints are never published. Good practice includes:

• Recording the version you read, such as arXiv v2, and its date.
• Checking whether a published version exists before submission, and citing it if so.
• Being cautious about comparing your results with numbers from preprints that have not been reviewed.

Step six: check code, data and hardware details

In engineering, a method is often only as reproducible as its code and experimental setup. When extracting information from each paper, record:

• Whether code is released, and where.
• Datasets or benchmarks used.
• Hardware platform, sensors and key parameters.
• Number of trials and how success was measured.

This information helps you choose fair baselines and identify gaps, such as methods that have only been tested in simulation or on a single robot.

Step seven: use citation chaining

Once you have a core set of highly relevant papers, follow their references backwards and their citations forwards. Open indexes make forward citation searching quick. Citation chaining often surfaces work that uses different terminology, which keyword searches miss, especially across disciplinary boundaries between robotics, computer vision and control.

Step eight: synthesise and write the related-work section

Organise related work by approach or problem dimension rather than chronologically or paper by paper. For each group:

• Summarise what the approaches have in common.
• Note their main strengths and limitations, ideally with evidence such as reported success rates or failure modes.
• Make clear how your work differs.

Every comparative claim, such as "prior methods require full state information" or "existing approaches have only been evaluated on cloth", should be checked against the sources. These are exactly the kinds of claims reviewers challenge.

Where evidence-linked AI assistants fit

AI research assistants can accelerate several steps: suggesting synonyms for search strings, surfacing papers across terminology boundaries, summarising methods and extracting experimental details into a table. The risk is that summaries become detached from their sources, and errors creep into the related-work section.

An evidence-linked assistant keeps each extracted claim tied to the passage it came from. You can ask "what hardware was used in the real-world experiments?" or "how many trials per condition?" and go straight to the paragraph or table that answers it. Yila is designed around this principle, searching real papers from OpenAlex, arXiv and other sources and attaching every statement to its evidence. Use such tools to navigate and organise, and always confirm key claims in the original text.

A minimal reproducibility checklist

1. A written scope statement and inclusion criteria.
2. A search log with databases, dates, strings and result counts.
3. Screening decisions with exclusion reasons.
4. A record of preprint versions and published versions.
5. An extraction table including code, data and hardware details.
6. A note of citation-chaining steps.
7. A date after which new work was not systematically searched.

Summary

A reproducible literature search turns related work from a last-minute scramble into a defensible part of an engineering paper. Define the scope, search the right venues, log every step, screen with explicit criteria, treat preprints and code with care, and verify every comparative claim. Tools that keep extracted evidence connected to its source, like the research workbench at https://yila.ai, make the process faster without loosening the link between what you write and what the literature actually says.