README.md

---
license: other
pretty_name: "RSS Papers — Robotics: Science and Systems (PDF archive)"
language:
  - en
tags:
  - robotics
  - papers
  - proceedings
  - academic
  - pdf
  - manipulation
task_categories:
  - other
size_categories:
  - n<1K
---

# RSS Papers — Robotics: Science and Systems (PDF archive)

A full-text **PDF archive of accepted papers** from the *Robotics: Science and Systems (RSS)* conference, organized by year, with per-paper metadata. Assembled as a convenience mirror of the publicly available official proceedings for personal research and archival use.

## Coverage

| Year | Edition | Papers |
|------|---------|-------:|
| 2026 | RSS XXII | 210 |
| 2025 | RSS XXI | 163 |
| 2024 | RSS XX | 134 |
| 2023 | RSS XIX | 112 |
| 2022 | RSS XVIII | 74 |
| 2021 | RSS XVII | 92 |
| 2020 | RSS XVI | 103 |
| **Total** | | **888** |

Years 2024–2026 additionally ship per-paper topic tags (`<year>/tags.json`) and an
aggregate topic breakdown (`<year>/analysis.json`) — see the analysis section below.

## Repository structure

```
README.md
<year>/
  papers.csv          # metadata for that year's papers
  p001_<title-slug>.pdf
  p002_<title-slug>.pdf
  ...
```

- One folder per year (`2024/`, `2025/`, `2026/`).
- PDF filename = zero-padded paper id + a slug of the title (e.g. `p007_a-systematic-study-of-data-modalities.pdf`).
- `<year>/papers.csv` columns:

| column | description |
|--------|-------------|
| `id` / `num` | paper number within the year |
| `session` | official RSS session (2026) / — |
| `title` | paper title |
| `authors` | author list |
| `filename` | PDF file name in this folder |
| `pdf_url` | original source URL the PDF was fetched from |
| `status` | crawl status (`downloaded`) |

## How to use

Download the whole dataset (or a single year) with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

# everything
path = snapshot_download("Ngseo/rss-papers", repo_type="dataset")

# just one year's PDFs + metadata
path = snapshot_download(
    "Ngseo/rss-papers", repo_type="dataset",
    allow_patterns=["2026/*"],
)
```

Read the metadata:

```python
import pandas as pd
df = pd.read_csv(f"{path}/2026/papers.csv")
print(df[["id", "session", "title"]].head())
```

## Topic landscape — RSS 2026 (content analysis)

Each 2026 paper was multi-label tagged into a fixed robotics taxonomy from its
title + abstract (LLM tagging, cross-checked against the official session labels
and unsupervised embedding clusters). Highlights:

**Application areas** (primary area of each paper)

| Rank | Area | Papers (primary) | Papers (tagged) |
|-----:|------|-----------------:|----------------:|
| 1 | Manipulation (incl. dexterous / grasping / mobile) | 92 | 125 |
| 2 | Mapping & Localization (SLAM) | 16 | 18 |
| 3 | Navigation | 15 | 39 |
| 4 | Humanoids | 14 | 23 |
| 5 | Multi-Robot & Swarm | 9 | 15 |
| … | Locomotion, HRI | 7 each | 22 / 17 |
| tail | Aerial, Soft, Tactile, Robot Design | ~5 each | — |
| minor | Autonomous Driving, Medical/Surgical, Field/Marine/Space | 3–4 each | — |

**Methods** (multi-label, most common first)

Imitation Learning (70) · Reinforcement Learning (56) · Perception/3D (53) ·
Foundation Models / LLM-VLM (50) · Self-Supervised Representation (48) ·
Motion & Task Planning (43) · Sim-to-Real (37) · Diffusion Policy (36) ·
Safety & Robustness (36) · MPC / Optimal Control (35) · Datasets & Benchmarks (33) ·
Vision-Language-Action (31) · Trajectory Optimization (27) · World Models (16).

**Input modalities:** Vision/RGB (116) · Proprioception (76) · Language (56) ·
Depth/Point-cloud (49) · Tactile/Force (27).

**Takeaways**

- **Manipulation is the dominant theme** — ~44% of papers have it as their primary
  area and ~60% touch it. The official session grouping *understates* this because
  manipulation work is spread across the *Imitation learning*, *VLA Models*,
  *World Models*, and *Datasets* sessions (RSS session slots are size-balanced for
  scheduling, so raw session counts flatten true topic prevalence).
- The **generative + foundation-model wave is large in aggregate**: Foundation
  Models (50) + Diffusion (36) + VLA (31) + World Models (16) together touch a big
  share of the program.
- **Under-represented** at RSS 2026: autonomous driving, medical/surgical robotics,
  soft robotics, and field/marine/space robotics (each only a handful of papers).

*(Analysis method: 3-lens triangulation — official sessions, LLM multi-label
tagging of title+abstract, and MiniLM embedding clustering.)*

## Source & provenance

- **2026:** paper list from <https://roboticsconference.org/program/papers/>, PDFs from <https://www.roboticsproceedings.org/rss22/>.
- **2024–2025:** <https://www.roboticsproceedings.org/> (`rss20`, `rss21`).

PDFs were fetched directly from the official proceedings; every file was validated
(PDF magic bytes + size) and cross-checked against the published paper index.

## License & usage

PDFs are © their respective authors and RSS. This is a convenience mirror of
publicly available proceedings for personal research and archival use; all rights
remain with the original authors/publisher. For the authoritative versions and
citation details, refer to the official RSS proceedings at
<https://www.roboticsproceedings.org/>.

If you use RSS papers in your work, cite the individual papers (and RSS) as listed
in the official proceedings.