This is not an official Google product.
This README describes the dataset used in the following publication: Hamrick, J. B., Ballard, A. J., Pascanu, R., Vinyals, O., Heess, N., & Battaglia, P. W. (2017). Metacontrol for Adaptive Imagination-Based Optimization. In Proceedings of the 5th International Conference on Learning Representations (ICLR 2017), available on openreview.
This repository contains the following CSV files, two for each of the five datasets corresponding to different numbers of planets:
Each row in these files corresponds to one scene, and each of these files contains the following columns:
- x_ship -- the x coordinate of the spaceship
- y_ship -- the y coordinate of the spaceship
- vx_ship -- the x velocity of the spaceship
- vy_ship -- the y velocity of the spaceship
- mass_ship -- the mass of the spaceship
- radius_ship -- the radius of the spaceship
- x_planetN -- the x coordinate of the Nth planet, where N is the identifier of the planet (e.g. 0, 1, 2, etc.)
- y_planetN -- the y coordinate of the Nth planet
- vx_planetN -- the x velocity of the Nth planet
- vy_planetN -- the y velocity of the Nth planet
- mass_planetN -- the mass of the Nth planet
- radius_planetN -- the radius of the Nth planet
- gravity -- the gravitational constant (G), which can be thought of as what is essentially a weight scale
- damping -- the damping coefficient
To simulate from these scenes, we computed the sum of forces acting on the spaceship from all the planets as well as a damping term (see Equation 9 in the paper) and the used Euler integration to simulate forward. When calculating forces, we accounted for one special case: If the spaceship was within the radius of a planet, the force exerted by the planet on the spaceship was set to a value as if the spaceship was just above the planet surface. We used an integrator step size of 0.05 and ran simulations forward for 12 steps.