backflip_AI

A humanoid learns backflips and other acrobatics by imitating mocap. The physics sim is C++ written from scratch, the learning is PPO in PyTorch, and they talk over UDP — 25 humanoids train in parallel.

Running it

You need GLFW, GLEW and OpenGL. Build the sim:

g++ src/mma.cpp -O2 -o mma -fopenmp -lglfw -lGLEW -lGL -lX11 -ldl -lpthread

Start it, then train against whatever clip is live:

./mma &
cd src && python train.py

To watch a trained policy instead, build with one humanoid rather than 25:

g++ src/mma.cpp -O2 -DNUM_ENVS=1 -o mma1 -fopenmp -lglfw -lGLEW -lGL -lX11 -ldl -lpthread
./mma1 &
cd src && cp ../models/backflip.pt POLICY_3D.pt && python test.py

r snaps to the start pose, r again runs from there, q quits. In the sim window b cycles how often it draws — training is much faster when it isn't rendering.

Switching clips

cd helpers
python use.py              # list what's available
python use.py backflip     # make that one live

It prints a DPHASE value — put it in test.py when you change clips.

Reverse RSI

Episodes normally start at a random point in the clip. That doesn't work for a backflip: the landing is only reachable if the takeoff and flight already went right, so a policy starting from frame 0 fails at the landing forever.

So the start window runs backwards instead. Episodes begin in the last 10% of the clip where all that's left is to land, and once that's solid the window widens back toward the takeoff. Every stage starts from something the previous stage already solved.

RC_ENABLE in src/train.py turns it on. Leave it off for walking and running, where every part of the cycle is reachable anyway and plain random starts are fine.

Layout

src/         sim, training, playback, and the keyframe animation editor
2D/          the earlier 2D version
animations/  reference clips
models/      trained policies, one per skill
helpers/     retarget mocap and check clips are physically possible

Clips are retargeted from DeepMimic, plus a few animated by hand.