ICRA Panel Tackles Paper Deluge in Robotics Publishing
At a recent ICRA panel titled "Surviving the Paper Deluge," chaired by Aude Billard, leading robotics researchers examined the exponential growth in robotics publications and debated how the field should adapt. The discussion ranged from practical uses of large language models to radical proposals for overhauling peer review.
Billard highlighted the explosion of papers in IEEE RAS venues, estimating roughly 70,000 papers containing the word "robotics" in 2025. She warned that robotics is integrative by nature, spanning perception, control, manipulation, learning, hardware, and deployment. If researchers only stay current within narrow silos, the field loses its core strength of connecting ideas across domains.
In a case study funded by IEEE RAS, Kunpeng Yao's team spent a year reading an entire subfield—learning from demonstration. They screened 2024 papers from IEEE Xplore and identified 347 relevant ones, but only 69 (about 20%) were judged notable. Notable papers offered new formulations, serious comparisons against strong state of the art, or convincing real-robot validation, while mere relabeling of existing approaches did not pass the bar.
On using large language models for literature review, Nadia Figueroa acknowledged that LLMs can improve search, retrieval, clustering, summaries, and conceptual maps, drastically reducing time. However, she described three levels of hallucination: first, fabricated references; second, real references but misstated content; and third, invented but plausible commonalities across papers. She warned that the current danger is not fake citations but fake understanding.
Greg Dudek criticized the practice of "salami slicing," where work is divided into the thinnest slices to inflate publication counts. The reward system encourages this, but it burdens readers and fragments knowledge. He advocated publishing fewer, more integrated papers, although he conceded there is no clean fix; comparing small procedural changes to building walls while a tsunami is approaching.
Renaud Detry argued that publication growth is partly due to more applied systems research, which should be evaluated differently from algorithmic papers. He proposed distinct tracks for fundamental science, applications, infrastructure, benchmarks, and technical correctness. Dongheui Lee added that visibility on arXiv or social media does not mean scientific value; editorial boards should promote strong papers that receive good reviews but lack famous-lab backing.
Shigeki Sugano pushed the most radical idea: abandoning traditional peer review altogether. He suggested uploading manuscripts to open archives with videos, code, and data, then having community evaluation under verified identities. Journals and conferences would certify high-value work instead of deciding what gets published. Panelists raised risks like popularity bias and unwanted noise, but the proposal highlighted the difficulty of scaling classical gatekeeping.
No definitive solution emerged, but several principles did. LLMs can offer mechanical support, but researchers must avoid outsourcing cognitive judgment. Readers should be selective and deep; institutions should reward quality over quantity; and different types of contributions need distinct recognition. Finally, as paper volumes grow, the robotics community must consider whether the flood will further silo the discipline or erode the shared foundation that keeps it integrative.