APPL is an agentic robot learning framework that connects policy learning with agent-based skill selection and composition through structural priors, enabling generalization from a few demonstrations to new scenes and task compositions.
A construction agent segments demonstrations into reusable skills, proposes structural priors, and trains and verifies a policy for each prior. A runtime agent reads interfaces that describe these priors and their applicability to select and compose policies for new task goals. The same structural prior shapes how a policy generalizes during learning and tells the runtime agent when to use it.
- Project page: APPL
- Code: GitHub
- Paper: arXiv:2609.35690
- Discussion: Hugging Face
This repository holds the static website only (index.html, static/). The rollout videos are Blender re-renders of logged ManiSkill episodes.
The page includes scholarly citation metadata, Schema.org ScholarlyArticle data, and a canonical URL. sitemap.xml lists the project page; llms.txt is an optional text guide to the same research content and sources, not a guarantee of AI indexing or citation.
Rollout descriptions are pre-rendered into index.html so crawlers can read their text without executing JavaScript. After editing static/data/tasks.json, run python3 scripts/render_task_panels.py and commit the generated HTML with the data. Run python3 scripts/render_task_panels.py --check to verify they match. No build step is needed when serving the checked-in website.
For discovery after deployment, submit https://agentics-robotics.github.io/APPL/sitemap.xml through the site's verified Google Search Console and Bing Webmaster Tools accounts. This project is hosted under /APPL/: crawler rules must be managed at https://agentics-robotics.github.io/robots.txt, in the domain-root site's repository. A /APPL/robots.txt file would not control crawlers. The domain-root file can also advertise this sitemap. Keep the project-page link in the arXiv record and other public research profiles up to date. Indexing and citations remain up to each search service.