Should a quadrotor’s airframe and its learned attitude controller be sized together, or one after the other?

What it found

A 15-minute inspection quadrotor with two airframe variables (arm length, prop diameter) and a three-gain roll policy found by CMA-ES, scored on held-out gusty step-tracking episodes. At 320 evaluations, joint CMA-ES reaches 5.25° RMS (the brute-force grid optimum is 5.25°), against 5.63° for joint random search, 5.52° for a frozen controller, 5.88° for bilevel search and 7.06° for design-first (lightest airframe, then tune). A controller tuned on one airframe and reused on others is 1° or more worse on 41% of cells. Gust level did not move the optimum. Stylised roll-axis model with assumed constants; the policy is linear, not deep RL; no flight software-in-the-loop.

A figure from its own output

Paper, code and results

• README and method
• Experiments and results
• Code
• Tests

More in Agentic co-design of low-cost robots

• Does the best quadrotor airframe stay best when motor strength, mass and motor lag are randomised?Does the best quadrotor airframe stay best under widened motor-strength, mass and lag randomisation? Mostly: top five unchanged (Spearman 0.75), but retuning gains helps the median airframe 0.91° (stylised).
• Should a quadrotor’s airframe and controller gains be searched jointly when the plant is uncertain?Should airframe and gains be searched jointly under an uncertain plant? Yes in this stylised model: joint CMA-ES at 960 evaluations (5.75° median held-out, 8 seeds) beats random search and a 4800-evaluation grid-then-tune best cell (6.00°).
• Is the quadrotor airframe ranking from one tuning run reproducible, or is it tuning noise?Is the airframe ranking from one tuning run reproducible? With wide-tuned gains yes (Spearman 0.99 over six replicates); with nominal-tuned gains it is noisier (0.84) and the earlier 0.17° regret was within noise (stylised).
• Does the ranking of printable soft-fruit gripper designs survive a shift in friction, fruit, placement error and print tolerance?Does the ranking of soft-fruit gripper designs survive a shift in friction, fruit, placement error and print tolerance? Yes in this model (Spearman 0.993 over 600 designs), but the nominal optimum loses 13% of its J and a robust optimum gains only about 0.015 on average; sim-to-sim, assumed ranges.
• Which grasp scenes can no printable two-finger gripper hold, and what limits the ceiling?Which grasp scenes can no printable two-finger gripper hold? 10 of 140 (7%), all spheres with the hand more than about 2 mm above their centre height; lowering only the hand holds all 10, so the 0.85 plateau in codesign-diagnose is scene geometry, not friction or force (stylised, found after the fact, one intervention).
• Does diagnostic feedback (a failure summary, full rollout logs) help a design loop find better printable grippers than the score alone?Does a failure summary or full rollout logs help a gripper design loop beyond the score alone? No: three scripted readers end at 0.84 to 0.86 best test success, like random search, CMA-ES and Bayesian optimisation (stylised, rule-based readers, not an LLM).

All 27 Agentic co-design of low-cost robots projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.