ICRA

\documentclass[conference]{IEEEtran}
\IEEEoverridecommandlockouts
\usepackage{cite}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}

\begin{document}

\title{Whole-Body Contact-Rich Manipulation via\\Differentiable Sim-to-Real Transfer}

\author{\IEEEauthorblockN{First Last}
\IEEEauthorblockA{\textit{University of Example}\\[email protected]}
\and
\IEEEauthorblockN{Jane Doe}
\IEEEauthorblockA{\textit{Example Research Labs}\\[email protected]}
\and
\IEEEauthorblockN{John Smith}
\IEEEauthorblockA{\textit{University of Example}\\[email protected]}}
\maketitle

\begin{abstract}
Contact-rich manipulation has been difficult to automate because of the
poor behavior of reinforcement learning in discontinuous dynamics. We
present a differentiable physics framework that enables direct
gradient-based training through contact. Real-robot experiments on a
7-DOF arm show a 4.1$\times$ improvement in success rate over standard
RL baselines across five contact-rich tasks.
\end{abstract}

\begin{IEEEkeywords}
contact-rich manipulation, differentiable simulation, sim-to-real
\end{IEEEkeywords}

\section{Introduction}
Contact-rich behaviors are ubiquitous in human manipulation yet remain
elusive for robots. The discontinuity of contact is not an obstacle if
one uses relaxed contact models that admit gradients.

\section{Related Work}
RL for manipulation, differentiable simulation, DeepMimic.

\section{Method}
We implement a smoothed contact model and train per-task policies by
taking gradients through rollouts:
\begin{equation}
  \pi_\theta(a \mid s) = \mathcal{N}(f_\theta(s), \Sigma).
\end{equation}

\subsection{Sim-to-Real Transfer}
We randomize physical parameters at training and adapt online via fast
system identification.

\section{Experiments}
\begin{table}[t]
\centering\small
\begin{tabular}{lcc}
\toprule
Task & PPO & \textbf{Ours} \\
\midrule
Push under clutter & 0.21 & \textbf{0.83} \\
Flip card          & 0.34 & \textbf{0.91} \\
Pour liquid        & 0.12 & \textbf{0.74} \\
\bottomrule
\end{tabular}
\caption{Success rate across 50 real-robot trials.}
\end{table}

\section{Discussion}
The dominant source of error in transfer is contact-stiffness mismatch;
our online adaptation closes the gap within 3 episodes.

\section{Conclusion}
Differentiable simulation plus online adaptation enables contact-rich
manipulation that was previously out of reach.

\bibliographystyle{IEEEtran}
\bibliography{refs}
\end{document}