Microduck Blind Basketball Balance in 3D Room (Recurrent LSTM `.onnx`)

Overview

Architecture: Blind proprioceptive LSTM(256) + ELU MLP(512, 256, 128) ([1, 61] observation + h_in/c_in [1, 1, 256] -> [1, 14] actions at 50 Hz, ~624k parameters). The deployed .onnx actor receives zero direct ball coordinates and infers rolling sphere contact dynamics purely from ankle/hip proprioception and body gyroscope history.
Honest MuJoCo Physics in Browser (duckspace.org): Free-rolling 6-DOF size-7 basketball (r = 0.12 m, mass = 0.62 kg, condim = 6, friction = 1.2 0.01 0.001), Rhoban BAM XL330 M6 voltage/back-EMF/load-dependent gearbox stiction actuator model (7.4 V, kt = 0.366, R = 2.81 Ohm), and 20 ms (4 physics substeps) bus command delay matching mjlab training.
Controls in 3D Room: Use WASD / Arrow keys or click on the apartment floor to steer the duck and roll the basketball across the room; drag the basketball with the mouse to test disturbance recovery, toggle the ball on/off in Room Settings (⚙️), or press Space to remount.

Licensed under the Apache License 2.0 (LICENSE and NOTICE included).

Original b11_6999 blind LSTM policy and training setup (Mjlab-Basketball-MicroDuck) by Hannes von Essen / Vottivott (Apache-2.0), built on pollen-robotics/microduck_rl and Rhoban/bam.
Live browser 50 Hz MuJoCo WebAssembly + ONNX Runtime Web 3D apartment integration by DuckSpace.org.

Video Preview