How to Start AI and Robotics Research (No Matter Your Background)

Starting AI and robotics research doesn’t require a specific degree level or a pre-built “research aptitude” resume. The practical unlock is to set expectations that match current background knowledge—and then follow a paper-reading strategy that turns curiosity into a repeatable workflow. The core message is that waiting for an arbitrary threshold usually backfires: the more people learn, the more they realize what they don’t know, so the goalposts keep moving. Instead, researchers should begin now, because early momentum opens doors and makes later learning faster.

The strategy begins with “appropriate expectations,” meaning a realistic view of what someone can gain right now from reading and doing research. Two archetypes illustrate the difference. Student one—new to the field, early in a bachelor’s program or just out of high school—may recognize some math and key terms, and understand the introduction’s language, but still miss how the work fits into the broader literature, what the contributions are, what the results mean, and how the results were generated. That’s not a failure; it’s a valid starting point. Student two—master’s level or later in a PhD, with literature review experience and at least one paper—reads with stronger context, understands the math formalism more deeply, and focuses on contributions, how results were produced, and how the work could be extended or improved. Even critique should be handled carefully: jumping straight to criticism can hide useful insights, so it’s often better to identify what’s strong first, then critique with awareness of potential bias.

With expectations aligned, the workflow shifts to “strategy,” starting by choosing a research area and learning how the field validates work. For robotics, major venues mentioned include IRA, IROS, and JRR, and the video emphasizes using Google Scholar to check where papers were published, how peer review signals value, and citation counts as a rough (imperfect) proxy for impact—especially noting that very new papers may have low citations simply due to recency.

Next comes the most actionable step: begin with a survey paper found via Google Scholar. Survey papers consolidate dozens of related works into a digestible map, offering multiple paths for deeper exploration. Access can be a constraint, so the transcript points to arXiv preprints as alternatives when journal or conference versions are paywalled, with the caveat that preprints may differ from final publications.

Finally, reading is done strategically rather than top-to-bottom. The method uses two handwritten “concept map” sheets. One sheet tracks keywords: as terms appear in the abstract, figures, and tables, unfamiliar words get written down without immediately looking them up. The second sheet captures figures and tables with one-sentence summaries in the reader’s own words. After this first pass, the reader uses the keyword list differently depending on experience level. Student one asks what background knowledge is needed—for example, if “convolutions” appear, the likely prerequisite is matrix multiplication and, more broadly, linear algebra. Student two asks how a term or architecture matters in context—why a specific design (like a D3QN architecture) is used, what it enables, and how it fits into the larger problem (such as terrain traversibility for mobile robots).

The payoff is faster comprehension, better paper selection, and a clearer sense of “what and why” behind each reading. The approach is intentionally slow at first, but writing in one’s own words and tying concepts to a bigger research question accelerates learning over time. Teaching what’s learned is recommended as a motivation and a diagnostic for understanding gaps.