Data explorer

Every joint of the nine example robots under five controllers, one row each: the drive's factory gains, the analytic start from the robot description, the same closed form fed the true model, what pidtuna's tuned run delivered, and the oracle, the best admissible controller a brute-force search found on the true plant. All five are scored on one shared held-out workload. Errors are in physical units (mrad, or mm on a linear axis); the relative cost is the tuner's own score, with factory gains at 1.

Search, sort and filter from the toolbar or the filters on the left, switch views with the tabs, and click a row to plot that joint's motion.

What the columns mean

• kp, kd: the joint's proportional and derivative gains, at the joint (N·m/rad and N·m·s/rad; N/m and N·s/m on a linear axis).
• ωn, ζ: the closed-loop natural frequency and damping ratio those gains give on the true plant (ωn = √(kp/J), ζ = kd ωn / 2kp).
• Stop p90: the 90th-percentile position error at the end of each hold. Tracking RMS: the RMS error while moving.
• Relative cost: stop p90 and tracking RMS, each over the factory gains', averaged: factory = 1, lower is better. vs oracle: that cost over the oracle's.
• Passes gates: no ringing, torque saturation and chatter within limits, tracking no worse than factory. A controller that fails a gate is never delivered, however low its cost: the Refused for column says why.

The data is /explorer/runs.json and the views are /explorer/views.base (an Obsidian Bases file), rendered by data-ui.