Duck walk

Night of 6 to 7 Nov. A walking policy trained in simulation for our duck: Waveshare ST3215-HS 12 V servos, about 2.5 kg, the heavier Microduck head, and no foot switches (they are not fitted). Trained on one rented RTX 4090, 150 M steps in 20 minutes, starting from the open-source Open Duck Mini v2 training code.

1. The trained walk (version A)

Forward, turn, forward while turning, side-step, backward, stop. The keys shown are the ones the real-duck runtime uses.

2. How it learned

Same duck model, same "walk forward" command, five points in training. The untrained policy outputs almost nothing, so it holds the standing pose; by 10 M steps it shuffles, by 43 M it walks.

3. Results in the simulator

PolicyNeeds on the duckForward in 8 sFalls
Stock open-source policy on our duck modelIMU + foot switches82 cmno
run1, trained for our duckIMU + foot switches107 cmno
A: no foot switches (main)servos + BNO055 IMU95 cmno
B: blind (backup)servos only74 cm, veers left about 40 cmno

4. Stress test of version A (10 s forward walk)

Change from what it trained onResult
Weight −20 % to +35 %walks, 112–137 cm
Servos 50 % softer to 2× stifferwalks, 103–123 cm
Servo damping ×0.5 to ×2walks, 90–100 cm
Extra control delay 20–40 mswalks
Extra control delay 60 msfalls after 6.7 s
Side shove 0.3 m/s every 2 swalks
Side shove 0.6 m/s every 2 sfalls

The datasheet estimates (weight, servo numbers) matter little. Delay is the real risk: the Radxa's loop must read the 14 servos and the IMU without adding more than about 40 ms. The policy itself takes 0.7 ms on the Radxa.

5. Status on the real duck

Everything above is simulation. Nothing has run on the real duck yet.