CoRL

Conference on Robot Learning paper using the official corl style. Two-column, focus on learning-based robotics.

Category

Conference

License

Free to use (MIT)

File

corl/main.tex

main.texRead-only preview
\documentclass{article}
\usepackage{corl_2024}  % swap to most recent corl style
\usepackage{graphicx}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}
\usepackage{natbib}
\usepackage[hidelinks]{hyperref}

\title{Skill Graphs: Compositional Imitation Learning\\
       for Long-Horizon Manipulation}

\author{
  First Last\thanks{University of Example, \texttt{[email protected]}} \and
  Jane Doe\thanks{Example Research Labs, \texttt{[email protected]}} \and
  John Smith\thanks{University of Example, \texttt{[email protected]}}
}

\begin{document}
\maketitle

\begin{abstract}
Long-horizon manipulation is unfriendly to flat imitation learning:
errors compound, and demonstrations are costly. We introduce Skill
Graphs, a compositional framework that factorizes tasks into reusable
skills with typed pre- and post-conditions. On a benchmark of 24
long-horizon tasks, Skill Graphs achieve 68\% success versus 24\% for
flat behavior cloning.
\end{abstract}

\keywords{imitation learning, manipulation, compositional methods}

\section{Introduction}
Robots that help in kitchens, workshops, and homes must stitch together
many short skills. Monolithic imitation scales poorly.

\section{Related Work}
Options framework, hierarchical RL, programmatic policies.

\section{Method}
A skill graph is a directed graph where nodes are skills and edges
encode typed pre- and post-conditions. Skills are learned from
50--100 demonstrations each; edges from fewer long-horizon demos.

\section{Experiments}
\begin{table}[t]
\centering
\begin{tabular}{lcc}
\toprule
Method & Short tasks & Long tasks \\
\midrule
BC                     & 0.82 & 0.24 \\
Diffusion              & 0.87 & 0.31 \\
\textbf{Skill Graphs}  & \textbf{0.89} & \textbf{0.68} \\
\bottomrule
\end{tabular}
\caption{Success rate on 24 real-robot tasks.}
\end{table}

\subsection{Ablations}
Removing typed conditions drops long-task success by 21 points.

\section{Conclusion}
Compositional imitation with typed conditions is a practical lever for
long-horizon manipulation.

\bibliographystyle{plainnat}
\bibliography{refs}
\end{document}
Bibby Mascot

PDF Preview

Create an account to compile and preview

CoRL LaTeX Template | Free Download & Preview - Bibby