AI ROBO STADIUM

Learn Physical AI by playing

Physical AI is AI that moves a body in the real world. Many legged robots learn to walk this way: they fall over again and again in simulation and improve through reinforcement learning.
AI ROBO STADIUM turns that reinforcement learning into a game. It is not an effect made to look like AI. PPO, a reinforcement learning algorithm also used in research, trains real neural networks inside your PC.

You teach. The AI learns to walk.

You decide what to reward and what to punish. How to drive the motors in the leg joints is something the AI figures out on its own by trial and error.
A robot that keeps falling at first learns to climb over steps and up slopes as you move it on to harder practice grounds.

Teach with methods from legged robot research

• Reward design: up to six levers, such as "Follow commands", "Stay upright", "Don't fall", "Save energy", "Move smoothly" and "Lift feet high", set what to reward and what to punish
  
• Domain randomization: vary floor friction, body mass and motor strength every episode, and see how it changes the way the robot falls in a real match
  
• Push training: shove the robots at random while they practice so they learn to keep their footing
  
• Terrain curriculum: move from flat floor to steps and slopes one practice ground at a time. Step height and slope angle rise automatically as training progresses
  
• Watch the learning curve, the fall rate and a breakdown of each reward term as training runs
  
• Take a model that is still in training and pilot it on the spot
  
• Save as many trained models as you like under your own names, and keep training any of them later

Real reinforcement learning, running on your PC

Training runs on your PC's CPU. It needs no GPU and no Python setup. 64 robots practice at once in a normal run, and 128 in an overnight run. You can keep racing and wrestling while it trains, and training continues even after you close the game.

Compete with the AI you raised

• Race: run an obstacle course of steps, slopes and a slalom against a rival robot in a single shared lane. Bumping is allowed, and a robot that cannot get up within 3 seconds loses
  
• Sumo: a one-on-one pushing match in the ring. Knock your opponent off balance with punch, lift and wedge arms
  
• You steer, the AI walks. You give the forward and turn commands, and the model you trained turns them into leg movements

New parts mean new training

Swap in a LiDAR, an IMU, foot contact sensors, wheels or upgraded motors. Bolting on a part does not make a robot stronger by itself. It has to learn to use the new loadout first.

Robots

• Gazelle: a quadruped
  
• Teeter: a biped with wheels at the end of its legs

The robots move in the game's physics simulation. Exporting a trained model to a real robot is not supported.
The AI in this game is a neural network that learns how to move the robot's joints through reinforcement learning. It does not generate text or images.