Uploaded December 2025 | Updated September 2026, 2 weeks ago
Rigid body moose learns to flip and land on its feet by a neural network which returns a dropping angle. The network uses a pre-generated training dataset of a couple of thousands of dropping angles and simply fit a function of rewards vs drop angle, where the reward is maximal for zero degree landing. Since this seems being a quite chaotic system, four hidden 512-neuron layers were used to approximate the wildly varying function from which the maximum reward was deduced. A simple grid search would have been better for this simple problem formulation, but additional complexity is probably more easily incorporated in this architecture.
A slightly less successful reinforcement learning agent tries afterwards predicting the same.
Song by @gpcbass called Alhambra.
Visuals with Python & FFmpeg
Rigid body moose learns to flip and land on its feet by a neural network which returns a dropping angle. The network uses a pre-generated training dataset of a couple of thousands of dropping angles and simply fit a function of rewards vs drop angle, where the reward is maximal for zero degree landing. Since this seems being a quite chaotic system, four hidden 512-neuron layers were used to approximate the wildly varying function from which the maximum reward was deduced. A simple grid search would have been better for this simple problem formulation, but additional complexity is probably more easily incorporated in this architecture.
A slightly less successful reinforcement learning agent tries afterwards predicting the same.
Song by @gpcbass called Alhambra.
Visuals with Python & FFmpeg
