A single-year review of learning-from-demonstration research found 347 relevant papers published in 2024, and only 69 of them, about 20 percent, made notable contributions. The result was presented at an ICRA panel on the robotics paper deluge, where researchers argued over how the field can cope with its publication volume and what should change in robotics peer review.
Panel chair Aude Billard pointed to rapid growth across major IEEE Robotics and Automation Society venues, on a roughly exponential curve that began around 2017. One estimate for 2025 puts the number of papers containing the word robotics at around 70,000.
What a year of reading found
Kunpeng Yao presented a case study funded by the IEEE RAS Science and Technology Watch Board. The team set out to read an entire subfield carefully over a single year. They searched IEEE Xplore for 2024 papers on learning from demonstration, screened the results manually and identified the 347 relevant papers.
Papers judged notable tended to offer new formulations, mechanisms or theoretical guarantees. They compared themselves against strong state-of-the-art baselines, showed clear gains in robustness, transfer or failure recovery, introduced new teaching or data collection modalities, and validated their claims on real robots. Papers were less convincing when they relabelled existing approaches or did not demonstrate what they claimed.
Yao's lesson was that notability should be judged against the state of the art, not against the newest vocabulary. Useful work, the panel noted, is sometimes hidden in less fashionable venues or written by less visible groups.
Billard argued that robotics cannot rely on narrow specialisation the way some fields can. A robot integrates perception, control, manipulation, learning, hardware, interaction, safety and deployment. If researchers can only follow narrow silos, the field loses its ability to connect ideas across domains.
LLMs and the problem of fake understanding
Nadia Figueroa said large language models can improve the mechanics of literature review, including search, retrieval, clustering, tables, summaries and rough conceptual maps. Work that once took weeks or months can sometimes be done in hours, which lowers the barrier for a new PhD student entering a field.
She described three orders of hallucination. The first is fake or incorrect references. The second is a real reference whose contents the model misstates. The third is the invention of plausible but false commonalities across papers. "The danger currently is not fake citations," Figueroa said. "It's fake understanding."
An LLM-based review can cite real papers and still misrepresent the field. If researchers hand over the comparison, judgment and doubt as well as the reading, the field may produce more text while developing less understanding.
Incentives, paper types and visibility
Greg Dudek described salami slicing, in which a body of work is divided into the thinnest possible slices and spread across workshops, conferences and journals. He said the system rewards it: students need papers to graduate, early-career researchers need them for jobs, and committees fall back on titles, venues, counts and indices. His preferred remedy is fewer, more integrated papers. He also said there is no fix if that means a return to a quieter world, and compared small procedural reforms to building better walls around a sandcastle while a tsunami approaches.
Renaud Detry said part of the growth reflects more applied systems work appearing in academic venues, which is not necessarily bad. The problem is that one evaluation process judges algorithms, engineered systems, benchmarks, datasets and industrial validation studies as if they were the same kind of paper. He suggested clearer tracks for fundamental science, applications, infrastructure, benchmarks and technical correctness.
Dongheui Lee addressed arXiv, open-source releases, project pages, videos, blogs and social media. These broaden access and help reproducibility, but visibility can reflect institutional prestige, networks, speed and self-promotion as much as scientific value. She proposed keeping expert peer review as a quality filter and using its signals better, with editorial boards and media teams promoting strongly reviewed papers from less famous labs.
Dropping the gatekeeper
Shigeki Sugano asked what would happen if peer review were abandoned altogether. In his model, legitimate robotics and AI manuscripts would go first to an open archive with videos, code and data. Community evaluation under verified identities would happen in public, and journals and conferences would certify high-value work rather than decide what gets to exist. "The question is not whether peer review is valuable," he said. "The question is whether it must remain the only gate to visibility."
Other panelists raised popularity bias, gaming, reputation effects and low-quality papers flooding the system as risks.
Analysis: what the panel does and does not establish
The most useful number from the session is the 20 percent figure, because it comes from a manual, criteria-based reading rather than from opinion. It suggests that the difficulty is less the total count than the cost of locating the minority of papers that move the frontier. That points toward better filtering and clearer paper categories as more practical targets than cutting volume, which Dudek himself described as unlikely.
The proposals also fit together unevenly. Detry's tracks and Lee's amplification of well-reviewed work both keep peer review in place and change what it signals. Sugano's model removes its gatekeeping role and moves the burden to public evaluation, which is exactly where the other panelists saw the risk of popularity bias. Lee's own point that visibility tracks prestige and self-promotion applies to open community review as much as to social media.
Several things remain unproven. The 20 percent result covers one subfield in one year, so it cannot be assumed to hold for other areas of robotics. The 70,000 figure is an estimate for papers containing a single word, not a count of distinct research contributions. None of the proposals was reported with test results, and the LLM benefits Figueroa described, hours instead of weeks, were not paired with any measured error rates for the second- and third-order hallucinations she warned about.
The practical implication is narrow but concrete. Any tool or policy aimed at the deluge should be judged on whether it helps readers find work that beats the state of the art and validates on real robots, not on whether it processes more papers. Until that is tested, the panel's agreement that no clean fix exists is the best-supported conclusion.