File size: 4,350 Bytes
590673f
1
2
<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><meta name="description" content="Run Dev Kumar's frozen UR7e SAC policy locally, compare a classical DLS controller and inspect a reproducible CPU benchmark."><title>UR7e Policy Lab · Dev Kumar</title><link rel="stylesheet" href="style.css"></head><body><main><header><a href="https://huggingface.co/devkumarhf" target="_blank" rel="noopener">DEV KUMAR / APPLIED ROBOTICS</a><span class="badge">LOCAL INFERENCE · NO API</span></header><section class="intro"><p class="eyebrow">MODEL → EXPERIMENT → EVIDENCE</p><h1>UR7e Policy Lab<span>.</span></h1><p>A frozen robot policy. A classical controller. The same trajectory.</p><div class="links"><a href="https://huggingface.co/devkumarhf/ur7e-sac-line-following" target="_blank" rel="noopener">Model &amp; weights ↗</a><a href="https://www.kaggle.com/code/devkumarkaggle/ur7e-sac-vs-dls-reproducible-policy-benchmark" target="_blank" rel="noopener">Experimental notebook ↗</a><a href="https://github.com/devkumar-projects/UR7e-Reinforcement-Learning-Suite" target="_blank" rel="noopener">Engineering source ↗</a></div></section><div class="lab"><section class="viewport"><div class="viewport-title"><span>6 AXES · KINEMATIC SIMULATION</span><span>Drag to orbit · Scroll to zoom</span></div><canvas id="scene" aria-label="Interactive 3D robot kinematic skeleton and target trajectory" tabindex="0"></canvas><div class="legend"><span class="target">Target</span><span class="actual">End-effector path</span><span>Live deterministic inference</span></div></section><aside><h2>Run the policy</h2><label>Controller<select id="controller"><option value="sac">Frozen SAC · 512 × 512</option><option value="dls">Damped least squares · fixed baseline</option></select></label><label>Trajectory family<select id="shape"><option value="segment">Straight segment</option><option value="bezier">Quadratic Bézier</option><option value="bezier3">Cubic Bézier</option><option value="sinus">Sinusoid</option><option value="arc">Circular arc</option></select></label><label>Prespecified example <span id="example-label">1 / 20</span><input id="example" type="range" min="0" max="19" value="0"></label><label>Sensor noise<select id="noise"><option value="0">None</option><option value="0.002">2 mm position / 2 mrad joints</option><option value="0.005">5 mm position / 5 mrad joints</option></select></label><div class="actions"><button id="run" disabled>Loading actor…</button><button id="reset">Reset</button></div><p id="status" role="status" aria-live="polite">Loading local model assets.</p><dl><div><dt>Trajectory progress</dt><dd id="progress">0%</dd></div><div><dt>Tracking RMS</dt><dd id="rms">—</dd></div><div><dt>Steps</dt><dd id="steps">0 / 800</dd></div><div><dt>Latest action time</dt><dd id="latency">—</dd></div></dl><p class="note">Preview runs entirely in JavaScript. Noise samples and numerical rounding differ from the Python reference benchmark.</p></aside></div><section class="evidence"><div><p class="eyebrow">REFERENCE EXPERIMENT</p><h2>600 CPU episodes.<br>Both sides of the result.</h2><p>100 paired trajectories × 2 controllers × 3 noise levels. No retraining. One frozen checkpoint.</p></div><div class="table-wrap"><table><caption>Measured Python evaluation · no added noise</caption><thead><tr><th>Controller</th><th>Strict completion</th><th>Tracking RMS*</th></tr></thead><tbody><tr><td>Frozen SAC</td><td>93 / 100</td><td>5.21 mm</td></tr><tr><td>Fixed DLS</td><td>90 / 100</td><td>2.38 mm</td></tr></tbody></table><p class="note">*Mean per-episode RMS over acquired trajectories, including failures. DLS tracks more accurately in this experiment. The completion difference does not establish statistical superiority. <a href="summary.json">All conditions and metadata ↗</a></p></div></section><footer><p>Position-only default kinematics. No dynamics, contact force, collision model or physical-robot validation. Original training seed history is unavailable.</p><p>Original project and policy: Dev Kumar. <a href="THIRD_PARTY_NOTICES.md">Robot-description attribution</a> · <a href="https://dev-kumar.com/" target="_blank" rel="noopener">Portfolio</a></p></footer></main><script type="module" src="app.js"></script></body></html>