Embodied AI & Robotics Papers Monitor (arXiv)

Get a clean, topic-tagged, deduplicated feed of new research papers on robot learning and embodied AI: vision-language-action (VLA) models, robotic manipulation, dexterous hands, humanoids, legged locomotion, sim-to-real, world models, navigation and more. Data comes straight from the official arXiv API. Run it once for a literature snapshot, or schedule it daily or weekly to get only the papers you haven't seen yet.

Why this Actor

Generic arXiv scrapers return whatever matches a search box. This one is pre-configured for the embodied-AI field:

• All of arXiv Robotics (cs.RO), plus the embodied-AI papers that are filed under AI, Computer Vision or Machine Learning (cs.AI, cs.CV, cs.LG) — those are only included when their title or abstract matches embodied-AI keywords (robot, embodied, VLA, humanoid, dexterous, sim-to-real, imitation learning, …), so you don't drown in general ML papers.
• Focused on embodied AI by default: every paper gets topic tags (VLA, Manipulation, Humanoid, Legged locomotion, Dexterous hands, Navigation, Imitation / robot learning, World models, Sim-to-real, Tactile sensing, …), and by default only papers with an embodied-AI tag are output. Classical robotics (drones, autonomous driving, SLAM, planning & control) is filtered out — clear Only these topics to get every robotics paper.
• Code links: GitHub / GitLab / Hugging Face links found in the abstract or comments, plus a has_code flag.
• Monitoring mode: scheduled runs output (and bill) only papers you haven't received before.

Use cases

• Robotics and embodied-AI teams: a weekly digest of new VLA, manipulation or humanoid papers.
• Researchers and students: keep up with a fast-moving field without reading the whole cs.RO listing.
• Investors and analysts: track research activity in humanoids, robot foundation models and world models.
• AI / RAG pipelines: clean title + abstract + metadata records, ready to embed or feed to an LLM.
• Alerts: schedule the Actor and connect it to Slack, email, Google Sheets or webhooks with Apify integrations.

How to use

1. Click Start with the defaults to get the last 7 days of embodied-AI papers.
2. Optional: narrow Only these topics (e.g. just VLA and Humanoid), clear it to get all robotics papers, or tick Only papers with a code link.
3. Results appear in the Output tab and can be downloaded as JSON, CSV, Excel or via API.
4. For monitoring, tick Only output papers not seen in previous runs and add a Schedule (e.g. daily).

Input

Field	Default	Description
categories	["cs.RO"]	Core arXiv categories: every paper is included
extraCategories	["cs.AI","cs.CV","cs.LG"]	Papers from these categories are included only if they match extraKeywords
extraKeywords	robot, embodied, VLA, humanoid, dexterous, sim-to-real, …	Embodied-AI phrases (title/abstract, any of them) for the extra categories
primaryCategoryOnly	true	Keep only papers whose main category is one of the above. Filtered papers are free
topicFilter	embodied-AI topics (VLA, Embodied agents, Manipulation, Dexterous hands, Humanoid, Legged locomotion, Navigation, Imitation learning, Robot learning, Diffusion / flow policies, World models, Sim-to-real, Tactile sensing)	Output only papers with at least one of these topic tags. Empty = all papers. Filtered papers are free
onlyWithCode	false	Output only papers with a GitHub/GitLab/Hugging Face link. Filtered papers are free
keywords	[]	Optional phrases that every paper must match
keywordsOperator / keywordFields	OR / title_abstract	Logic and fields for keywords
lastNDays	7	Date window; ignored if dateFrom/dateTo set; 0 = no date filter
dateFrom / dateTo	–	YYYY-MM-DD (UTC, inclusive)
dateField	submittedDate	or lastUpdatedDate
maxRecords	100	Max papers output (and charged) per run
sortBy / sortOrder	submittedDate / descending	also lastUpdatedDate, relevance
onlyNew	false	Skip papers seen in previous runs (free)
includeNewVersions	false	With onlyNew, output a paper again when a higher version appears
seenStoreName	embodied-ai-papers-seen	Named key-value store for seen IDs; use a different name per monitor
maxScannedRecords	3000	Cap on arXiv results scanned when client-side filters are on
contactEmail	–	Optional override of the contact in the User-Agent sent to arXiv

Example: a daily monitor for VLA and humanoid papers that come with code.

{
  "topicFilter": ["VLA", "Humanoid"],
  "onlyWithCode": true,
  "lastNDays": 3,
  "maxRecords": 200,
  "onlyNew": true,
  "seenStoreName": "vla-humanoid-monitor"
}

Output example

{
  "source": "arxiv",
  "arxiv_id": "2610.12345",
  "version": 1,
  "title": "Smooth Action Chunking Via Movement-Primitive Representations for Imitation Learning",
  "authors": [{"name": "…", "affiliations": []}],
  "abstract": "…",
  "categories": ["cs.LG", "cs.RO"],
  "primary_category": "cs.LG",
  "published": "2026-10-28T17:59:59Z",
  "updated": "2026-10-28T17:59:59Z",
  "pdf_url": "https://arxiv.org/pdf/2610.12345v1",
  "abs_url": "https://arxiv.org/abs/2610.12345v1",
  "doi": null,
  "journal_ref": null,
  "comments": "Project page: https://…",
  "topics": ["VLA", "Manipulation", "Imitation learning"],
  "code_urls": ["https://github.com/…"],
  "project_urls": ["https://…"],
  "has_code": true,
  "is_new": true
}

Topic tags are assigned by transparent keyword rules on the title and abstract (no AI model), so they are fast and predictable but not perfect: a paper can have several tags or none. Each run also saves a RUN_SUMMARY record (output / scanned / filtered / skipped counts, stop reason, the exact arXiv query) to its default key-value store.

Pricing

$2 per 1,000 papers ($0.002 per output paper, event paper-record). Papers skipped as already seen, or removed by the category, topic or code filters, are free. The run stops cleanly at your maximum charge limit, so you never pay more than you set. Example: the default 7-day window has roughly 230 embodied-AI papers (default maxRecords is 100, i.e. $0.20 per run; about $0.47 for all of them); a daily monitor with onlyNew typically outputs 30–50 new papers (about $0.08).

Tips

• arXiv announces new submissions with a delay (typically ~1 business day, none on weekends), so the most recent day of a window is usually still empty. For daily schedules use lastNDays ≥ 3 with onlyNew on. The overlap costs nothing because already-seen papers are not charged.
• Typical volume (checked Nov 2026, 7 days): ~233 papers with the default topic filter (197 with primary category cs.RO, 36 from cs.CV, cs.AI and cs.LG); ~324 with the topic filter cleared.
• Want everything in robotics, including drones, driving and SLAM? Clear Only these topics.
• Each arXiv page (≤100 papers) takes ≥3 s because of the rate limit, so 1,000 papers take about 30 s or more.
• Abstracts may contain LaTeX ($...$); it is passed through unchanged.

FAQ

Is this allowed? Yes. The Actor uses arXiv's official, public API within its published rate limits and only collects public metadata (titles, abstracts, authors, links). It does not download PDFs.

Why are some general ML papers included? Papers from cs.AI, cs.CV and cs.LG are included when they mention robots or embodied-AI terms and get an embodied-AI topic tag. Remove a word from Embodied-AI keywords, or select fewer topics, to make the feed stricter.

Why is a robotics paper missing? By default only papers with an embodied-AI topic tag are output. Clear Only these topics to get every paper.

Can I run several monitors? Yes. Give each one its own seenStoreName so their "seen" lists don't mix.

Data & attribution

Metadata is retrieved from the arXiv API. Thank you to arXiv for use of its open access interoperability. Paper content remains under each paper's own license; this Actor does not download or redistribute PDFs.