Five-Link Inverted Pendulum

I built a physics simulation of a cart with five connected pendulums and trained a controller to swing them up and keep them balanced. The final approach uses imitation learning from optimized expert trajectories, with behavioral cloning and DAgger.

The included controller succeeded in 512/512 evaluation runs. Success means all five poles stay within 10° of upright and below 0.5 rad/s for 2 seconds without the cart leaving the track.

Run it

Requires Python 3.11+ and the included model checkpoint (runs/imitation_complete/best.pt).

pip install -r requirements.txt
python play.py

Controls: R reset · D disturb a joint · Space pause · Esc quit.

To evaluate the model:

python evaluate.py --episodes 512 --seed 91000

To train a new model (requires experts/library.pt):

python train.py --expert experts/library.pt --output runs/imitation_new

* ChatGPT was used to generate the test and expert code.