Embedded Files
Arne Troch - Kevin Mets - Siegfried Mercelis
arne.troch@uantwerpen.be
Control Systems
Control Systems
Datasets can be downloaded from: https://huggingface.co/datasets/Atroch/ControlEnvs
pendulum_dataset_episodes.mp4Pendulum
Pendulum
acrobot_dataset_episodes.mp4Acrobot
Acrobot
cartpole_dataset_episodes.mp4CartPole
CartPole
reacher_dataset_episodes.mp4Reacher
Reacher
cart2pole_dataset_episodes.mp4Cart2Pole
Cart2Pole
cart3pole_dataset_episodes.mp4Cart3Pole
Cart3Pole
acro3bot_dataset_episodes.mp4Acro3bot
Acro3bot
Comparing ACHGN++ with Latent Neural ODE
Comparing ACHGN++ with Latent Neural ODE
Pendulum_ACHGNvsLatNODE_Zero.mp4β-ACHGN++ vs Latent Neural ODE
β-ACHGN++ vs Latent Neural ODE
Pendulum without Control
Pendulum without Control
→ energy conservation
→ energy conservation
β-ACHGN++ conserves energy while LatNODE reduces in phase space volume over time
Pendulum_ACHGNvsLatNODE_Control.mp4β-ACHGN++ vs Latent Neural ODE
β-ACHGN++ vs Latent Neural ODE
Pendulum with Control
Pendulum with Control
→ energy variations
→ energy variations
β-ACHGN++ closely matches energy while LatNODE underestimates the energy present
Regularizing ACHGN++
Regularizing ACHGN++
The rate of change of the system energy should equal the power input from the external forces
→ Adding this to training objectives leads to better long term predictions!
→ Adding this to training objectives leads to better long term predictions!
Steps before diverging from ground truth:
Interpreting latent space of ACHGN++
Interpreting latent space of ACHGN++
Trained model uses only a single latent coordinate!
Trained model uses only a single latent coordinate!
This coordinate behaves just like the real canonical coordinates from the ground truth!
This coordinate behaves just like the real canonical coordinates from the ground truth!
Page updated
Google Sites
Report abuse