4 ms·

https://ehmorris.com/lander/ https://ehmorris.com/lander/

hathawsh 3y ago
Oh my goodness this is gold. I've never had so much fun failing.
ViewTrick1002 3y ago
Felt completely intuitive. All hours playing Kerbal Space Program re-loading the last save after crashing into the surface paying off....
- blharr 3y ago
  If it's too easy, try doing a couple flips for more points
  - hoten 3y ago
    If it's too hard, just do many dozens of flips as you try to race away from the surface.
    - ehmorris 3y ago
      The real challenge is landing after you get really really high up
super256 3y ago
There is this experiment in OpenAI Gym to land a spacecraft via reinforcement learning. Is reinforcement learning actually working for this, and are such models deployed to real rockets, instead of "hardcoded" math? https://gymnasium.farama.org/environments/box2d/lunar_lander/ https://gymnasium.farama.org/environments/box2d/lunar_lander...
- ehmorris 3y ago
  Someone made a neat autopilot for an older version of the game actually! https://szhu.github.io/lunar-lander-autopilot/ https://szhu.github.io/lunar-lander-autopilot/
- wizzwizz4 3y ago
  Reinforcement learning does work for this, but it's brittle. You almost always end up with a strategy that exploits inaccuracies in the physics simulation – from the perspective of the RL algorithm, useful quirks of the laws of physics – and so doesn't transfer to reality very well, if at all. Its behaviour outside the conditions observed in training is not guaranteed (or even expected) to be sensible, and even its behaviour within the training conditions is often hard to characterise. An algorithm designed by people who know what they're doing is usually better. More effort, yes, but rockets are a lot of effort! We can afford to pay the cost for a reliable landing system.
nick7376182 3y ago
If you go fast enough sideways, the lander goes off screen and you can fall past the ground. Then you can flip for unlimited reward. Unfortunately, you can't then crash to see a final score! Best I could do was a 970 point crash.

Oh my goodness this is gold. I've never had so much fun failing.

Felt completely intuitive. All hours playing Kerbal Space Program re-loading the last save after crashing into the surface paying off....

blharr 3y ago
If it's too easy, try doing a couple flips for more points
- hoten 3y ago
  If it's too hard, just do many dozens of flips as you try to race away from the surface.
  - ehmorris 3y ago
    The real challenge is landing after you get really really high up

super256 3y ago

There is this experiment in OpenAI Gym to land a spacecraft via reinforcement learning. Is reinforcement learning actually working for this, and are such models deployed to real rockets, instead of "hardcoded" math? https://gymnasium.farama.org/environments/box2d/lunar_lander/ https://gymnasium.farama.org/environments/box2d/lunar_lander...

ehmorris 3y ago
Someone made a neat autopilot for an older version of the game actually! https://szhu.github.io/lunar-lander-autopilot/ https://szhu.github.io/lunar-lander-autopilot/
wizzwizz4 3y ago
Reinforcement learning does work for this, but it's brittle. You almost always end up with a strategy that exploits inaccuracies in the physics simulation – from the perspective of the RL algorithm, useful quirks of the laws of physics – and so doesn't transfer to reality very well, if at all. Its behaviour outside the conditions observed in training is not guaranteed (or even expected) to be sensible, and even its behaviour within the training conditions is often hard to characterise. An algorithm designed by people who know what they're doing is usually better. More effort, yes, but rockets are a lot of effort! We can afford to pay the cost for a reliable landing system.

nick7376182 3y ago

If you go fast enough sideways, the lander goes off screen and you can fall past the ground. Then you can flip for unlimited reward. Unfortunately, you can't then crash to see a final score! Best I could do was a 970 point crash.