The Publishing Crisis Nobody in Robotics Wants to Talk About

Seventy thousand papers. That's the staggering estimate for robotics publications in 2025, according to a recent ICRA panel discussion on the state of academic publishing. To put that in perspective, if you dedicated every waking hour to reading papers — no research, no teaching, no actual robotics work — you still couldn't keep up with a fraction of that output.

This isn't a growth problem. It's a crisis of coherence.

The panel, titled "Surviving the paper deluge," addressed what everyone in the field knows but rarely says out loud: we've created a system that rewards quantity over communication. The incentive structures in academic robotics push researchers to publish relentlessly, fragment their work across multiple papers, and prioritize incremental contributions that can secure conference slots. The result is a field that's producing more knowledge than it can possibly synthesize.

The ironic twist? The proposed solution involves the very technology driving much of this research explosion. Large language models are now being pitched as tools to accelerate literature review, helping researchers wade through mountains of papers faster than humanly possible. On the surface, this seems sensible. Who wouldn't want an AI assistant to summarize thousands of papers and identify relevant work?

But the panel raised a crucial concern: LLMs risk introducing "sophisticated hallucinations" into the literature review process. We're not just talking about made-up citations or misattributed findings — though those are problems. We're talking about AI systems that can generate plausible-sounding summaries of research that subtly misrepresent methodology, overstate conclusions, or miss critical nuances that only a human expert would catch.

Using AI to manage an AI-driven publication explosion feels like fighting fire with gasoline.

The real issue isn't whether LLMs can technically help researchers read faster. It's that we're treating a systemic problem with a technological band-aid. The robotics community hasn't seriously reckoned with why we're publishing at this volume in the first place, or whether the current system serves its core purpose: advancing the field through shared knowledge.

Some fields have started experimenting with radical alternatives. The Distill journal, before it ceased publication, focused on exceptionally clear explanations rather than novel results. Some AI research groups are moving toward living documents that get updated rather than spawning endless sequels. Open-source robotics projects often communicate progress through GitHub repositories and working demos rather than paper mills.

But academic robotics remains locked in a publish-or-perish cycle that's reaching an absurd endpoint. When the field produces more papers annually than any individual could read in a lifetime, we've lost the plot. Publications cease to be communication and become mere credentials — checkboxes for tenure committees and grant applications.

The ICRA panel deserves credit for surfacing this issue publicly. But until the incentive structures change — until universities, funding agencies, and conferences reward clarity and impact over volume — we'll continue drowning in our own output. Adding AI lifeguards won't save us. We need to stop flooding the pool.