Learning Where it Matters: Autonomous Data Collection and Offline Deep Reinforcement Learning for Local Contact-Rich Manipulation

Project page (supplementary video material) for an anonymous submission under double-blind review.

We autonomously collect robot data where it matters — at the contact-rich segment of a task — and learn from it with offline deep reinforcement learning. The result is a policy that is robust and generalizable, yet very precise. We modularly integrated learned local policies with motion planning to solve long-horizon tasks.

Hosting (GitHub Pages)

This repository is a static site. Enable GitHub Pages with:

• Settings → Pages → Build and deployment → Source: Deploy from a branch
• Branch: main / / (root)

The site is then served at the repository's Pages URL. A .nojekyll file is included so the static/ directory is served verbatim.

Structure

index.html                  # the page
paper.pdf                   # the paper
static/
  css/index.css             # styles
  js/index.js               # scroll reveals + viewport-based video autoplay
  images/method.png         # method figure
  images/posters/*.jpg      # video poster frames
  videos/*.mp4              # web-optimized, faststart-enabled clips

All videos are sped up and web-optimized (H.264, +faststart for progressive playback).