💬 NLP

Charting the Growth of Social-Physical HRI (spHRI): A Systematic Review Pipeline Augmented by Small Language Models

This paper presents a systematic review of social-physical human-robot interaction (spHRI) that demonstrates how locally running small language models (SLMs) can effectively augment human expert screening by significantly accelerating the process and identifying relevant papers that reviewers might otherwise miss.

Original authors: Mayumi Mohan, Ju-Hung Chen, Alexis E. Block

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mayumi Mohan, Ju-Hung Chen, Alexis E. Block

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Needle in a Growing Haystack

Imagine you are trying to find every single book in a massive library that talks about "robots touching people in a friendly way." This is a field called social-physical human-robot interaction (spHRI).

The authors of this paper noticed two big problems:

The Library is Exploding: The number of books (research papers) on this topic is growing incredibly fast. It's like a snowball rolling down a hill, getting bigger and bigger every year.
The Library is Messy: Everyone uses different words to describe the same thing. One scientist calls it "robotic hugging," another calls it "affective touch," and a third calls it "social contact." This makes it very hard to find all the relevant books using a standard search.

Traditionally, to find these books, you need a team of human experts to read the titles and short summaries (abstracts) of thousands of papers. This is like hiring a team of librarians to read every single book cover to decide if it belongs on your shelf. It takes a long time, costs a lot of money, and is exhausting.

The Experiment: Can a "Pocket-Sized" AI Help?

The researchers wanted to know if they could use Small Language Models (SLMs) to help.

Think of a Large Language Model (LLM) (like the famous ones you might know) as a super-smart, PhD-level librarian who lives in a giant, expensive data center. They are brilliant but require a lot of electricity and money to run.

The Small Language Models (SLMs) used in this study are like smart pocket calculators. They aren't as powerful as the PhD librarian, but they are tiny, run on a regular desktop computer, and are free to use.

The team asked: Can these tiny, local "pocket calculators" help us find the right books faster, even if they aren't as smart as the human experts?

How They Did It

The Human Team: Two human reviewers went through a massive list of papers and marked which ones were relevant (kept) and which were not (discarded).
The AI Team: They ran four different tiny AI models on the same list. These models were told: "Read the title and summary. Is this about robots touching people socially? Say 'Yes' or 'No'."
The Safety Net: To make sure the AI didn't accidentally keep too many bad papers, they used a "unanimity rule." A paper was only flagged for a second look if all four tiny AIs agreed it was a "Yes."

What They Found

The results were a mix of "not quite ready to replace humans" and "actually very useful helpers."

1. The AI isn't the Boss (Yet)
The human reviewers were still much better at deciding what was relevant. The tiny AIs made more mistakes than the humans. If you asked the AI to do the whole job alone, it would miss important papers or keep too many irrelevant ones.

Analogy: It's like asking a smart calculator to write a novel. It can do math fast, but it can't write a story as well as a human author.

2. The AI is a Speed Demon
While the humans took 30 to 60 seconds to read one paper, the tiny AIs did it in less than a second. They were thousands of times faster.

Analogy: The humans are walking through the library; the AI is a high-speed train zooming through the aisles.

3. The "Safety Net" Success
This is the most important finding. When the AI flagged papers that the humans had discarded, the humans went back and re-read them.

Result: The AI found 39 papers that the humans had missed.
Impact: These 39 papers made up 10.3% of the final list of important studies.
Analogy: Imagine the humans were fishing for gold. They caught a lot of gold, but the AI was a second net dragged behind the boat. The second net caught 10% of the gold that the first net missed. Without that second net, they would have lost a significant chunk of their treasure.

The Conclusion

The paper concludes that these tiny, local AI models shouldn't replace human experts. Instead, they should be used as a safety net.

Think of it like a spell-checker. You wouldn't let a spell-checker write your essay for you, but you would definitely want it to scan your work to catch mistakes you missed.

By using these small, local models as a "second pair of eyes," researchers can:

Scan massive amounts of data very quickly.
Catch important papers that human fatigue caused them to miss.
Do this without needing expensive cloud servers or huge energy bills.

The authors argue that because the field of robot-touch research is growing so fast, we need this kind of "AI-assisted" help to keep our knowledge organized and complete.

Technical Summary: Charting the Growth of Social-Physical HRI (spHRI)

Problem Statement
Social-physical human-robot interaction (spHRI), defined as physical interactions with robots where social intent is the primary goal, has experienced rapid growth across robotics, human-computer interaction (HCI), and haptics. However, the field suffers from fragmented terminology, disciplinary silos, and inconsistent methodologies, which hinder systematic knowledge accumulation. While large-scale systematic reviews are necessary to map this landscape, the sheer volume of literature (over 130,000 candidate publications in this study) creates a bottleneck for human reviewers. Traditional screening is time-consuming, and while Large Language Models (LLMs) offer a potential solution, they often require cloud access, substantial computational resources, and raise ethical concerns regarding carbon footprints and data opacity. Consequently, there is a need for accessible, environmentally responsible, and scalable screening tools that do not rely on expensive infrastructure.

Methodology
The authors conducted a systematic literature review following a preregistered PRISMA workflow to investigate the growth of spHRI and evaluate the utility of Small Language Models (SLMs) as a screening aid.

Data Collection: Records were retrieved from PubMed, IEEE Xplore, Scopus, and the ACM Digital Library using a refined Boolean search query covering robots, touch/haptics, and social/affective concepts. After removing duplicates, the initial dataset comprised approximately 182,000 records, from which 32,000 were screened by two independent human researchers.
SLM Implementation: Four open-source, lightweight SLMs (≤ 1.5 billion parameters) were evaluated: Llama3.2 (1B), Gemma3 (1B), Qwen3 (0.6B), and DeepSeek-R1 (1.5B). These models were deployed locally on a consumer-grade desktop (Intel i5-13600K, NVIDIA RTX4070) using the Ollama framework.
Ensemble Strategy: The models were tested individually and as a "unanimity ensemble." In the ensemble approach, a paper was flagged for human re-screening only if all four SLMs predicted it as "relevant" (include). This conservative rule was designed to minimize false positives and keep the manual re-screening workload manageable.
Evaluation Metrics: Performance was measured against human reviewer labels using accuracy, recall, precision, specificity, false positive rate (FPR), and F1 score. Statistical inference utilized paired nonparametric bootstrapping (10,000 resamples) with Holm–Bonferroni correction.

Key Contributions

Systematic Mapping of spHRI: The study provides the first comprehensive systematic review spanning HRI, HCI, haptics, and robotics, clarifying the conceptual foundations and methodological landscape of spHRI.
SLM Benchmarking in HRI: The paper benchmarks consumer-grade SLMs against human reviewers for title and abstract screening in a specific, nuanced domain (spHRI), moving beyond the typical biomedical or software engineering benchmarks found in existing literature.
Accessible Review Pipeline: It demonstrates a practical workflow for integrating locally deployed, low-parameter models into systematic reviews, offering a sustainable alternative to cloud-based LLMs.

Results

Growth Trajectory (H1 Supported): Analysis of publication counts from 1992 to 2025 reveals a sustained, nonlinear growth in spHRI research, with pronounced acceleration in the last decade. A Generalized Additive Model (GAM) fitted to the data accounted for 96.9% of the variance, confirming spHRI as an emergent subfield.
Performance vs. Humans (H2 Rejected): No single SLM or the combined ensemble matched human reviewer performance. All models performed significantly worse than humans across all metrics ( $p < 0.001$ ). However, Llama3.2 achieved the best standalone agreement with human labels (Specificity = 0.9366) while screening papers in 0.21 seconds each, orders of magnitude faster than the 30–60 seconds required by humans.
Utility as a Safety Net (H3 Supported): The unanimity ensemble successfully identified 39 relevant papers that were missed by the initial human screening. These 39 papers represented 10.29% of the final relevant dataset (379 papers total), exceeding the study's predefined threshold of 5%.
Trade-offs: The results highlight a precision-recall tradeoff. Models like Gemma achieved high recall (0.9794) but with extreme over-inclusion (FPR = 0.9185). Conversely, the ensemble approach was highly selective (Precision = 0.2202) with a very low FPR (0.00161), effectively acting as a "safety net" to catch missed papers without overwhelming reviewers with false positives.

Significance and Claims
The paper argues that SLMs should not be viewed as replacements for expert human reviewers but as augmentative tools that make large-scale literature reviews accessible and sustainable. The authors claim that:

SLMs can serve as reliable "second-pass filters" to recover missed relevant studies, thereby improving the comprehensiveness of systematic reviews.
Local deployment of small models democratizes the systematic review process, allowing research labs without access to high-performance cloud computing or expensive API credits to conduct rigorous, large-scale syntheses.
The field of spHRI has matured enough to be recognized as a distinct subfield, but its fragmented nature requires the kind of systematic synthesis that human effort alone struggles to scale, making AI-assisted workflows essential for future research consolidation.

The authors acknowledge limitations, including the use of binary (no-confidence) SLM outputs and a single-pass human screening process, which precludes inter-rater reliability analysis. They position this work as a Late-Breaking Report (LBR) that establishes a foundation for refining prompts, exploring alternative ensembles, and expanding screening to full datasets in future work.