Research notes sometimes deserve to look like a paper — title, abstract,
sections, tables, bibliography — not just inline math. This site can now
compile full LaTeX documents at build time and render them in the page as
real content (ar5iv-style), in the same spirit as the .ipynb notebook cells.
Two tags — pandoc turns the LaTeX into page content, pdfLaTeX compiles the same source to a PDF for download:
| Tag | What it does |
|---|---|
{% tex_cell %} |
Renders the inline LaTeX in the tag body |
{% tex_file "path.tex" %} |
Renders an external .tex file |
Inline cell
The tag body is wrapped in a small article preamble for you (unless it
already contains \documentclass), so you can just write the paper:
A Short Note on Why I Like Robots
2026-08-04
Abstract
This note exists purely to show that inline LaTeX compiles to a real, down-loadable PDF.
Why robots
Robots are fun. Table 1 summarises the evidence.
| Robot | Fun | Upside |
|---|---|---|
| Wheeled | High | Medium |
| Legged | Higher | High |
Conclusion
More robots, more fun.
.tex source
\title{A Short Note on Why I Like Robots}
\author{Shivaram Kumar Jagannathan}
\date{\today}
\maketitle
\begin{abstract}
This note exists purely to show that inline LaTeX compiles to a real,
down-loadable PDF.
\end{abstract}
\section{Why robots}
Robots are fun. Table~\ref{tab:robots} summarises the evidence.
\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
Robot & Fun & Upside \\
\midrule
Wheeled & High & Medium \\
Legged & Higher & High \\
\bottomrule
\end{tabular}
\caption{Funness by platform.}
\label{tab:robots}
\end{table}
\section{Conclusion}
More robots, more fun.The cell above rendered a full title block, abstract, table, and caption as selectable page content — no build config, no helper scripts.
Full paper from a file
For a real manuscript, keep the .tex under assets/papers/ and embed it
with tex_file. The file carries its own \documentclass so it is compiled
verbatim — bibliography via \begin{thebibliography} and all:
Planning Is Downstream: A Perception–Action View of Robot Intelligence
2026-08-04
Abstract
Classical robot planning reasons over a fixed symbolic model of the world. We argue that this abstraction is the root cause of brittleness in the face of uncertainty. Drawing on cognitive neuroscience and the Free Energy Principle, we sketch a complementary view in which planning, perception, and action form a single loop: every action is simultaneously an attempt to achieve a goal and an opportunity to improve the agent’s world model. We formalise the loop, show how it subsumes classical planning as a special case, and point at the open problems that remain.
Introduction
Task and Motion Planning (TAMP) decomposes manipulation problems into symbolic planning over predicates and geometric reasoning over motion [1]. The symbolic layer is expressed in PDDL [2], where actions are described by preconditions and effects. This representation is computationally attractive but brittle: when a precondition fails because of observation error or an incomplete model, the planner re-searches the same fixed action space and typically finds no better answer.
Biological agents behave differently. When an action fails, they do not search harder in the same space; they act to improve their understanding of the world. They change their viewpoint, probe uncertain objects, and gather sensory evidence. We call this the perception–action loop.
A Minimal Model
Let \(m \in \mathcal{M}\) denote the agent’s internal world model, \(a \in \mathcal{A}\) an action, and \(o \in \mathcal{O}\) an observation. Classical planning fixes \(m\) and searches over \(a\) to reach a goal \(g\):
\[\begin{equation} a^{*} = \operatorname*{arg\,max}_{a} \; \Pr(g \mid a, m). \label{eq:planning} \end{equation}\]
The perception–action view instead treats \(m\) as a control variable too. Each action yields a reward for task progress and information about the world, quantified by the mutual information between the observation and the model:
\[\begin{equation} a^{*} = \operatorname*{arg\,max}_{a} \; \underbrace{\Pr(g \mid a, m)}_{\text{task}} \;+\; \beta \, I(m; o \mid a). \label{eq:paloop} \end{equation}\]
When \(\beta = 0\), Equation (2) reduces to classical planning, Equation (1). The agent that actively gathers information is exactly the one described by the Free Energy Principle [3]: perception and action jointly minimise variational free energy, so that acting reduces surprise about future observations.
Empirical Preview
As a sanity check we compare a fixed-model planner against the active version in a small gridworld with partial observability, over \(10^4\) episodes.
| Agent | Success \(\uparrow\) | Steps \(\downarrow\) | Model error \(\downarrow\) |
|---|---|---|---|
| Fixed-model planner | 0.58 | 14.3 | 0.41 |
| Active planner (\(\beta=0.5\)) | 0.87 | 9.1 | 0.12 |
Table 1 supports the claim that information-gathering actions are not a detour but the mechanism by which new affordances become available. Results are preliminary; a full study is deferred to future work.
Conclusion
Language, planning, and reasoning are downstream capabilities. They should emerge from an embodied, continually evolving world representation rather than serve as its foundation. The primary objective of robot learning is therefore not to imitate behaviour, but to acquire robust grounded models of the world through interaction — and that model can only be built through the continual cycle of perception and action.
Acknowledgements
The author thanks the readers of the blog post this note is distilled from.
References
[1]C. R. Garrett, R. Chitnis, R. Holladay, B. Kim, T. Silver, L. P. Kaelbling, and T. Lozano-Pérez. Integrated task and motion planning. Annual Review of Control, Robotics, and Autonomous Systems, 2021.
[2]D. McDermott, M. Ghallab, A. Howe, C. Knoblock, A. Ram, M. Veloso, D. Weld, and D. Wilkins. PDDL — the planning domain definition language. Technical Report CVC TR-98-003, Yale Center for Computational Vision and Control, 1998.
[3]K. Friston. The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7):293–301, 2009.
.tex source
% ---------------------------------------------------------------------
% Demo: a self-contained academic paper. Compile it in place with:
% {% tex_file "/assets/papers/example-paper.tex" %}
% It carries its own \documentclass, so it is compiled verbatim.
% ---------------------------------------------------------------------
\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}
\usepackage{microtype}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{enumitem}
\usepackage[colorlinks=true,urlcolor=blue,linkcolor=black,citecolor=blue]{hyperref}
\setlength{\parskip}{0.4em}
\title{Planning Is Downstream: A Perception--Action View of Robot Intelligence}
\author{Shivaram Kumar Jagannathan\\
\small{Nanyang Technological University, Singapore}\\
\small{\texttt{[email protected]}}}
\date{\today}
\begin{document}
\maketitle
\begin{abstract}
Classical robot planning reasons over a fixed symbolic model of the world.
We argue that this abstraction is the root cause of brittleness in the face of
uncertainty. Drawing on cognitive neuroscience and the Free Energy Principle,
we sketch a complementary view in which planning, perception, and action form
a single loop: every action is simultaneously an attempt to achieve a goal and
an opportunity to improve the agent's world model. We formalise the loop, show
how it subsumes classical planning as a special case, and point at the open
problems that remain.
\end{abstract}
\section{Introduction}
\label{sec:intro}
Task and Motion Planning (TAMP) decomposes manipulation problems into symbolic
planning over predicates and geometric reasoning over motion
\cite{garrett2021tamp}. The symbolic layer is expressed in PDDL
\cite{mcdermott1998pddl}, where actions are described by preconditions and
effects. This representation is computationally attractive but brittle: when a
precondition fails because of observation error or an incomplete model, the
planner re-searches the same fixed action space and typically finds no better
answer.
Biological agents behave differently. When an action fails, they do not search
harder in the same space; they act to \emph{improve} their understanding of the
world. They change their viewpoint, probe uncertain objects, and gather sensory
evidence. We call this the \emph{perception--action loop}.
\section{A Minimal Model}
\label{sec:model}
Let $m \in \mathcal{M}$ denote the agent's internal world model, $a \in
\mathcal{A}$ an action, and $o \in \mathcal{O}$ an observation. Classical
planning fixes $m$ and searches over $a$ to reach a goal $g$:
\begin{equation}
a^{*} = \operatorname*{arg\,max}_{a} \; \Pr(g \mid a, m).
\label{eq:planning}
\end{equation}
The perception--action view instead treats $m$ as a control variable too.
Each action yields a reward for task progress \emph{and} information about the
world, quantified by the mutual information between the observation and the
model:
\begin{equation}
a^{*} = \operatorname*{arg\,max}_{a} \; \underbrace{\Pr(g \mid a, m)}_{\text{task}} \;+\; \beta \, I(m; o \mid a).
\label{eq:paloop}
\end{equation}
When $\beta = 0$, Equation~\eqref{eq:paloop} reduces to classical planning,
Equation~\eqref{eq:planning}. The agent that actively gathers information is
exactly the one described by the Free Energy Principle
\cite{friston2009free}: perception and action jointly minimise variational
free energy, so that \emph{acting} reduces surprise about future observations.
\section{Empirical Preview}
\label{sec:preview}
As a sanity check we compare a fixed-model planner against the active version
in a small gridworld with partial observability, over $10^4$ episodes.
\begin{table}[h]
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
Agent & Success $\uparrow$ & Steps $\downarrow$ & Model error $\downarrow$ \\
\midrule
Fixed-model planner & 0.58 & 14.3 & 0.41 \\
Active planner ($\beta=0.5$) & \textbf{0.87} & 9.1 & 0.12 \\
\bottomrule
\end{tabular}
\caption{The active planner succeeds more often and in fewer steps, because it
spends early actions reducing model error.}
\label{tab:results}
\end{table}
Table~\ref{tab:results} supports the claim that information-gathering actions
are not a detour but the mechanism by which new affordances become available.
Results are preliminary; a full study is deferred to future work.
\section{Conclusion}
\label{sec:conclusion}
Language, planning, and reasoning are downstream capabilities. They should
emerge from an embodied, continually evolving world representation rather than
serve as its foundation. The primary objective of robot learning is therefore
not to imitate behaviour, but to acquire robust grounded models of the world
through interaction --- and that model can only be built through the continual
cycle of perception and action.
\section*{Acknowledgements}
The author thanks the readers of the blog post this note is distilled from.
\begin{thebibliography}{9}
\bibitem{garrett2021tamp}
C.~R. Garrett, R.~Chitnis, R.~Holladay, B.~Kim, T.~Silver, L.~P. Kaelbling,
and T.~Lozano-P\'erez.
\newblock Integrated task and motion planning.
\newblock \emph{Annual Review of Control, Robotics, and Autonomous Systems},
2021.
\bibitem{mcdermott1998pddl}
D.~McDermott, M.~Ghallab, A.~Howe, C.~Knoblock, A.~Ram, M.~Veloso,
D.~Weld, and D.~Wilkins.
\newblock PDDL --- the planning domain definition language.
\newblock Technical Report CVC TR-98-003, Yale Center for Computational
Vision and Control, 1998.
\bibitem{friston2009free}
K.~Friston.
\newblock The free-energy principle: a rough guide to the brain?
\newblock \emph{Trends in Cognitive Sciences}, 13(7):293--301, 2009.
\end{thebibliography}
\end{document}
Note the numbered equation cross-references (\eqref) and citations
([1], [2], [3]) that resolve against the manual bibliography — a full
research paper, rendered in place in Computer Modern, the classic LaTeX font.
Every cell gets
- Real page content — pandoc converts the LaTeX to HTML: selectable,
searchable, mobile-friendly text in Computer Modern (self-hosted Latin
Modern). Math stays as TeX and is typeset by the site’s MathJax with
amsequation numbering. - A
.pdfdownload — the same source is compiled with pdfLaTeX and staged next to the page for download. - A collapsible
.texsource — syntax-highlighted, so readers can copy the source directly.
Errors surface instead of crashing the build
Broken sources don’t break the build — they render an error card with the log, so you can fix the source. If pandoc can still convert the LaTeX but pdfLaTeX fails (say, a missing package), the page content renders anyway and the toolbar shows a “PDF failed” warning instead of the download link. This cell is deliberately broken:
broken-cellSyntax reference
{% tex_cell label=my-paper %} <!-- optional name for the .pdf -->
\title{...}\maketitle
\begin{abstract}...\end{abstract}
\section{...}...
{% endtex_cell %}
{% tex_file "/assets/papers/example-paper.tex" %}
tex_cellauto-wraps in\documentclass[11pt]{article}+geometry,amsmath,hyperref,booktabs,enumitem, etc. Write a full\documentclassyourself to opt out.- Compiled artifacts are cached in
.jekyll-cache/tex/and only recompiled when the source changes. - Requires
pandoc(HTML),pdflatex(TeX Live, PDF), andpdfinfo(poppler, page count) on the machine that builds the site. Binary paths can be overridden with thePANDOC/PDFLATEX/PDFINFOenv vars.