| Binding | Description |
|---|---|
| Right Mouse Click (Drag) | Rotates the camera in the direction of the mouse drag. |
| Left Mouse Click | Sets the goal point to the location on the terrain where you clicked. |
| W | Move forward. |
| S | Move backward. |
| A | Move left. |
| D | Move right. |
| E | Move up. |
| Q | Move down. |
| R | Restart the episode. Note that the terrain does not change even if you are in random mode. Deformations to the terrain are preserved as well. |
| 1 |
Terrain Mode: RANDOM
Restart with a random terrain shape (no perlin noise).
|
| 2 |
Terrain Mode: RANDOM_NOISE
Restart with a random terrain shape (with perlin noise).
|
| 3 |
Terrain Mode: GAUSSIAN_SLOPE
Restart with a gaussian slope shaped terrain.
|
| 4 |
Terrain Mode: POINTY_GAUSSIAN_SLOPE
Restart with a pointy gaussian slope shaped terrain.
|
| 5 |
Terrain Mode: TRAPEZOIDAL_HILL
Restart with a trapezoidal hill shaped terrain.
|
Unfortunately, MLAgents doesn't support GPU inference for WebGL builds.
The excavator model was originally trained with a larger neural network and yielded good results, but the
model was too large to run smoothly on the CPU.
The current model was trained with a smaller model. Although it can run more smoothly on the CPU, the
performance isn't as good as that of the larger model.
If you would like to try out the larger model as well, refer to excav_demo. When using the larger model, it is recommended that you download the build corresponding to your specific target platform.
For more information about WebGL performance considerations, refer to the unity documentation .