Hugging Face's logo

Spaces:
SabaPivot
/
repro-understanding-behavior-cloning-with-action-quantization
Running

App Files Community
repro-understanding-behavior-cloning-with-action-quantization / poster.html
SabaPivot's picture
SabaPivot
Upgrade reproduction with executable evidence, hosted rerun, poster, and audit bundle
a8c50be verified
Raw
History Blame Contribute Delete
15.8 kB
<!doctype html>
<!-- DESIGN DIRECTION (Step 2.5, locked)
 concept: control-room trace — six response panels separate statistical error from deployed behavior
 axis1 layout: 3-column evidence grid; two navigable claim regions per column
 axis2 canvas: warm near-white, flat
 axis3 palette: topic-derived accent plus contrast register
 axis4 type: sans display + serif body
 axis5 density: balanced
 axis6 card: white plate, thin frame, restrained shadow
 axis7 heading: numbered roundel + short rule
 axis8 masthead: left editorial band; compact audit footer
 devices: metric bars, rate ladder, structure matrix
 hero moment: the six evidence regions as one repeated visual system
 soft-clash notes: none
-->
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>Understanding Behavior Cloning with Action Quantization — reproduction poster</title>
<style>
@page { size: 60in 36in; margin: 0; }
:root {
 /* ===== DESIGN TOKENS ===== */
 --accent: #355F39;
 --accent-deep: #1F4023;
 --accent-soft: #E2EEE3;
 --accent-pale: #F0F6F0;
 --emph: #8F2437;
 --emph-soft: #F7E5E8;
 --ink: #201C26;
 --ink-soft: #554E5E;
 --ink-muted: #77707E;
 --paper: #F5F1EC;
 --card: #FFFFFF;
 --line: #C8DAC9;
 --line-strong: #87A98A;
 --white: #FFFFFF;
 --screen: #27232D;
 --shadow: rgba(52, 34, 82, 0.09);
 --u: 1.6px;
 --fs-1: calc(7 * var(--u));
 --fs-2: calc(9 * var(--u));
 --fs-3: calc(10 * var(--u));
 --fs-4: calc(11 * var(--u));
 --fs-5: calc(13 * var(--u));
 --fs-6: calc(15 * var(--u));
 --fs-7: calc(19 * var(--u));
 --fs-8: calc(24 * var(--u));
 --fs-9: calc(25 * var(--u));
 --font-sans: "Inter", "Helvetica Neue", Arial, sans-serif;
 --font-serif: "Charter", "Source Serif Pro", Georgia, serif;
 /* ===== END DESIGN TOKENS ===== */
}
* { box-sizing: border-box; margin: 0; padding: 0; }
html, body {
 background: var(--screen);
 color: var(--ink);
 font-family: var(--font-serif);
}
.poster {
 width: calc(1524 * var(--u));
 height: calc(914 * var(--u));
 margin: 20px auto;
 padding: calc(11 * var(--u)) calc(15 * var(--u));
 background: var(--paper);
 display: grid;
 grid-template-columns: minmax(0, 1fr);
 grid-template-rows: auto minmax(0, 1fr) auto;
 gap: calc(10 * var(--u));
 overflow: hidden;
 box-shadow: 0 0 calc(20 * var(--u)) var(--shadow);
}
.masthead {
 display: grid;
 grid-template-columns: minmax(0, 1fr) auto;
 gap: calc(14 * var(--u));
 align-items: end;
 padding: calc(4 * var(--u)) calc(5 * var(--u)) calc(7 * var(--u));
 border-bottom: calc(2 * var(--u)) solid var(--accent);
}
.eyebrow {
 color: var(--emph);
 font-family: var(--font-sans);
 font-size: var(--fs-2);
 font-weight: 800;
 letter-spacing: calc(.6 * var(--u));
 text-transform: uppercase;
}
h1 {
 color: var(--accent-deep);
 font-family: var(--font-sans);
 font-size: var(--fs-9);
 line-height: 1.03;
 letter-spacing: calc(-.2 * var(--u));
 margin-top: calc(2 * var(--u));
 max-width: calc(1000 * var(--u));
 text-wrap: balance;
}
.dek {
 color: var(--ink-soft);
 font-family: var(--font-sans);
 font-size: var(--fs-5);
 line-height: 1.25;
 margin-top: calc(3 * var(--u));
 text-wrap: pretty;
}
.stamp {
 min-width: calc(185 * var(--u));
 padding: calc(5 * var(--u)) calc(7 * var(--u));
 border: calc(.6 * var(--u)) solid var(--line-strong);
 background: var(--card);
 color: var(--accent-deep);
 font-family: var(--font-sans);
 font-size: var(--fs-3);
 line-height: 1.35;
 text-align: center;
}
.stamp strong {
 display: block;
 color: var(--emph);
 font-size: var(--fs-6);
}
.body {
 min-height: 0;
 display: grid;
 grid-template-columns: repeat(3, minmax(0, 1fr));
 gap: calc(9 * var(--u));
}
.column {
 min-height: 0;
 display: flex;
 flex-direction: column;
}
.plate {
 height: 100%;
 min-height: 0;
 display: grid;
 grid-template-rows: repeat(2, minmax(0, 1fr));
 background: var(--card);
 border: calc(.6 * var(--u)) solid var(--line);
 border-top: calc(3 * var(--u)) solid var(--accent);
 border-radius: calc(3 * var(--u));
 box-shadow: 0 calc(2 * var(--u)) calc(7 * var(--u)) var(--shadow);
 overflow: hidden;
}
.claim {
 min-height: 0;
 padding: calc(7 * var(--u)) calc(8 * var(--u));
 display: grid;
 grid-template-rows: auto auto auto minmax(0, 1fr) auto;
 gap: calc(3 * var(--u));
}
.claim + .claim { border-top: calc(.6 * var(--u)) solid var(--line); }
.claim-head {
 display: flex;
 align-items: center;
 gap: calc(4 * var(--u));
}
.num {
 width: calc(16 * var(--u));
 height: calc(16 * var(--u));
 flex: 0 0 auto;
 border-radius: 50%;
 background: var(--accent);
 color: var(--white);
 display: grid;
 place-items: center;
 font-family: var(--font-sans);
 font-size: var(--fs-4);
 font-weight: 800;
}
h2 {
 color: var(--accent-deep);
 font-family: var(--font-sans);
 font-size: var(--fs-7);
 line-height: 1.08;
 text-wrap: balance;
}
.verdict {
 color: var(--emph);
 font-family: var(--font-sans);
 font-size: var(--fs-2);
 font-weight: 800;
 letter-spacing: calc(.25 * var(--u));
 text-transform: uppercase;
}
.copy {
 color: var(--ink-soft);
 font-size: var(--fs-4);
 line-height: 1.32;
 text-wrap: pretty;
}
.copy strong { color: var(--ink); }
.visual {
 min-height: calc(64 * var(--u));
 align-self: stretch;
 display: grid;
 align-items: end;
 gap: calc(3 * var(--u));
 padding: calc(5 * var(--u)) calc(6 * var(--u));
 border-radius: calc(2 * var(--u));
 background: var(--accent-pale);
 border: calc(.5 * var(--u)) solid var(--accent-soft);
}
.mesh { grid-template-columns: repeat(5, 1fr); }
.bar {
 min-height: calc(12 * var(--u));
 background: var(--accent);
 border-radius: calc(1 * var(--u)) calc(1 * var(--u)) 0 0;
}
.mesh .bar:nth-child(1) { height: 28%; }
.mesh .bar:nth-child(2) { height: 48%; }
.mesh .bar:nth-child(3) { height: 67%; }
.mesh .bar:nth-child(4) { height: 84%; }
.mesh .bar:nth-child(5) { height: 98%; background: var(--emph); }
.curves {
 grid-template-columns: repeat(8, 1fr);
 background: var(--emph-soft);
 border-color: var(--emph-soft);
}
.curves .bar:nth-child(1) { height: 27%; }
.curves .bar:nth-child(2) { height: 36%; }
.curves .bar:nth-child(3) { height: 45%; }
.curves .bar:nth-child(4) { height: 55%; }
.curves .bar:nth-child(5) { height: 65%; }
.curves .bar:nth-child(6) { height: 75%; }
.curves .bar:nth-child(7) { height: 86%; }
.curves .bar:nth-child(8) { height: 98%; }
.matrix {
 grid-template-columns: repeat(5, 1fr);
 grid-template-rows: repeat(5, 1fr);
 align-items: stretch;
}
.cell {
 min-height: calc(9 * var(--u));
 background: var(--accent-soft);
 border-radius: calc(1 * var(--u));
}
.cell:nth-child(6n+1) { background: var(--accent); }
.cell:nth-child(3n+2) { background: var(--line-strong); }
.cell:nth-child(4n) { background: var(--emph-soft); }
.inverse {
 grid-template-columns: repeat(6, 1fr);
}
.inverse .bar:nth-child(1) { height: 98%; background: var(--emph); }
.inverse .bar:nth-child(2) { height: 78%; }
.inverse .bar:nth-child(3) { height: 61%; }
.inverse .bar:nth-child(4) { height: 49%; }
.inverse .bar:nth-child(5) { height: 40%; }
.inverse .bar:nth-child(6) { height: 34%; }
.ladder {
 grid-template-columns: 1fr;
 align-items: center;
 gap: calc(4 * var(--u));
}
.rung {
 height: calc(9 * var(--u));
 background: var(--accent);
 border-radius: calc(5 * var(--u));
}
.rung:nth-child(1) { width: 96%; }
.rung:nth-child(2) { width: 82%; }
.rung:nth-child(3) { width: 70%; }
.rung:nth-child(4) { width: 61%; }
.rung:nth-child(5) { width: 54%; background: var(--emph); }
.metric-row {
 display: grid;
 grid-template-columns: repeat(3, minmax(0, 1fr));
 gap: calc(2 * var(--u));
}
.metric {
 padding: calc(3 * var(--u));
 background: var(--accent-soft);
 color: var(--accent-deep);
 font-family: var(--font-sans);
 font-size: var(--fs-2);
 line-height: 1.25;
 text-align: center;
}
.metric strong {
 display: block;
 color: var(--emph);
 font-size: var(--fs-6);
 line-height: 1.05;
}
.control {
 padding-top: calc(3 * var(--u));
 border-top: calc(.5 * var(--u)) solid var(--line);
 color: var(--ink-muted);
 font-family: var(--font-sans);
 font-size: var(--fs-2);
 line-height: 1.25;
 text-wrap: pretty;
}
.control strong { color: var(--accent-deep); }
.footer {
 display: grid;
 grid-template-columns: minmax(0, 1fr) auto;
 gap: calc(10 * var(--u));
 align-items: center;
 padding: calc(3 * var(--u)) calc(5 * var(--u));
 border-top: calc(.6 * var(--u)) solid var(--line-strong);
 color: var(--ink-muted);
 font-family: var(--font-sans);
 font-size: var(--fs-2);
}
.footer strong { color: var(--accent-deep); }
.footer .right { color: var(--accent); font-weight: 700; }
@media print {
 :root { --u: 1mm; }
 html, body {
 width: 100%;
 height: 100%;
 margin: 0;
 padding: 0;
 background: var(--white);
 overflow: hidden;
 }
 .poster {
 width: 60in;
 height: 36in;
 margin: 0;
 box-shadow: none;
 page-break-after: avoid;
 }
}
</style>
</head>
<body>
<main class="poster" data-measure-role="poster">
 <header class="masthead" data-measure-role="header">
 <div>
 <div class="eyebrow">Independent reproduction · six registered claims</div>
 <h1>Understanding Behavior Cloning with Action Quantization</h1>
 <p class="dek">A scaled numerical audit of quantization floors, closed-loop instability, and rollout correction.</p>
 </div>
 <div class="stamp"><strong>4 · 2</strong>source exact · registered wording misstated</div>
 </header>
<section class="body" data-measure-role="body">
 <div class="column" data-measure-role="column">
 <article class="plate" data-measure-role="card">
 <section class="claim" data-logbook-target="claim-1-behavior-cloning-with-quantized-actions-and-log-loss">
 <div class="claim-head"><span class="num">1</span><h2>Quantized estimation rate</h2></div>
 <div class="verdict">Supported · scaled audit</div>
 <p class="copy">MSE falls with sample size at slope −0.938 (R² = 0.997) before reaching the quantization floor.</p>
 <div class="visual mesh" aria-label="Bars summarizing the audited trend"><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span></div>
 <div class="control"><strong>Control:</strong> Unquantized sample means remove the floor.</div>
 </section>
 <section class="claim" data-logbook-target="claim-2-under-probabilistic-incremental-input-to-state-stability-p">
 <div class="claim-head"><span class="num">2</span><h2>Stable versus unstable rollout</h2></div>
 <div class="verdict">Supported · premise stress test</div>
 <p class="copy">Stable dynamics show polynomial-like horizon growth; ρ = 1.3 produces exponential growth with semilog slope 0.267.</p>
 <div class="visual matrix" aria-label="Matrix cells summarizing a structural certificate"><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span><span class="cell"></span></div>
 <div class="control"><strong>Control:</strong> The unstable run deliberately violates the stability premise.</div>
 </section>
 </article>
 </div>
 <div class="column" data-measure-role="column">
 <article class="plate" data-measure-role="card">
 <section class="claim" data-logbook-target="claim-3-theorem-6-shows-that-without-a-smoothness-assumption">
 <div class="claim-head"><span class="num">3</span><h2>One-step error can mislead</h2></div>
 <div class="verdict">Supported · counterexample mechanism</div>
 <p class="copy">One-step error falls from 0.72 to 0.045 as εq shrinks, while deployed regret remains fixed at 0.63 per horizon.</p>
 <div class="visual ladder" aria-label="Horizontal rungs summarizing the rate"><span class="rung"></span><span class="rung"></span><span class="rung"></span><span class="rung"></span><span class="rung"></span></div>
 <div class="control"><strong>Control:</strong> A smooth quantizer makes deployed error shrink.</div>
 </section>
 <section class="claim" data-logbook-target="claim-4-theorem-7-proves-that-model-based-data-augmentation">
 <div class="claim-head"><span class="num">4</span><h2>Rollout augmentation</h2></div>
 <div class="verdict">Misstated · registered citation or scope</div>
 <p class="copy">Auxiliary rollout correction restores linear εq scaling (slope 1.000) and reaches a 21.91× improvement at the finest grid.</p>
 <div class="visual inverse" aria-label="Descending bars summarizing the controlled comparison"><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span></div>
 <div class="control"><strong>Control:</strong> Without augmentation the deployed curve stays constant.</div>
 </section>
 </article>
 </div>
 <div class="column" data-measure-role="column">
 <article class="plate" data-measure-role="card">
 <section class="claim" data-logbook-target="claim-5-information-theoretic-lower-bounds-theorems-8-9-establish">
 <div class="claim-head"><span class="num">5</span><h2>Deterministic and stochastic rates</h2></div>
 <div class="verdict">Supported · formula audit</div>
 <p class="copy">The registered lower-bound forms cleanly separate 1/n deterministic behavior from 1/√n stochastic behavior at fixed quantization error.</p>
 <div class="visual curves" aria-label="Bars summarizing the parameter sweep"><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span></div>
 <div class="control"><strong>Control:</strong> Setting εq = 0 isolates the statistical terms.</div>
 </section>
 <section class="claim" data-logbook-target="claim-6-empirically-binning-quantizers-are-shown-to-preserve-policy">
 <div class="claim-head"><span class="num">6</span><h2>Hard-policy discontinuity</h2></div>
 <div class="verdict">Misstated · registered citation or scope</div>
 <p class="copy">Adjacent-grid TV change is 5.93e−5 for a soft binned policy but remains exactly 1 for the learned hard rule.</p>
 <div class="visual mesh" aria-label="Bars summarizing the independent audit"><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span><span class="bar"></span></div>
 <div class="control"><strong>Control:</strong> Grid refinement shrinks the soft jump, not the hard jump.</div>
 </section>
 </article>
 </div>
 </section>
<footer class="footer" data-measure-role="footer">
 <div><strong>Scope:</strong> independent deterministic/scaled numerical audit · seed 20260726 · theorem checks do not replace proofs.</div>
 <div class="right">OpenReview 9uENnRAcSl · arXiv 2603.20538</div>
 </footer>
</main>
</body>
</html>