Running the simulation

This section will describe how to run the physics simulation and collect data. The code in this section is reflected in examples/basic/simple_physics.py.

Start by displaying a simple scene with a cube above a plane, viewed by a camera:

import simulate as sm
scene = sm.Scene(engine="Unity")
scene += sm.LightSun()
scene += sm.Box(
    name="floor",
    position=[0, 0, 0],
    bounds=[-10, 10, -0.1, 0, -10, 10],
    material=sm.Material.GRAY75,
)
scene += sm.Box(
    name="cube",
    position=[0, 3, 0],
    scaling=[1, 1, 1],
    material=sm.Material.GRAY50,
    with_rigid_body=True,
)
scene += sm.Camera(name="camera", position=[0, 2, -10])
scene.show()

# Prevent auto-closing when running locally
input("Press enter to continue...")

Note that we use the Unity engine backend, which supports physics simulation, as well as specify with_rigid_body=True on the cube, to enable forces like gravity.

Next, run the simulation for 30 timesteps:

for i in range(60):
    event = scene.step()

You should see the cube falling onto the plane.

step() tells the backend to step the simulation forward, and allows keyword arguments to be passed, allowing a wide variety of customizable behavior. The backend then returns a dictionary of data as an event. By default, this dictionary contains nodes and frames.

nodes is a dictionary containing all assets in the scene and their physical parameters such as position, rotation, and velocity. Try graphing the height of the cube as it falls:

import numpy as np
import matplotlib.pyplot as plt
plt.ion()
_, ax1 = plt.subplots(1, 1)
heights = []
for i in range(60):
    event = scene.step()
    height = event["nodes"]["cube"]["position"][1]
    heights.append(height)
    ax1.clear()
    ax1.set_xlim([0, 60])
    ax1.set_ylim([0, 3])
    ax1.plot(np.arange(len(heights)), heights)
    plt.pause(0.1)

frames is a dictionary containing the rendering from each camera. Try modifying the code to display these frames in matplotlib:

plt.ion()
_, ax1 = plt.subplots(1, 1)
for i in range(60):
    event = scene.step()
    im = np.array(event["frames"]["camera"], dtype=np.uint8).transpose(1, 2, 0)
    ax1.clear()
    ax1.imshow(im)
    plt.pause(0.1)

🤗 Simulate is highly customizable. If you aren’t interested in returning this data, you can modify the scene configuration prior to calling show() to disable it:

scene.config.return_nodes = False
scene.config.return_frames = False
scene.show()

For advanced use, you can extend this functionality using plugins.

In this library, we include an extensive plugin for reinforcement learning. If you are using 🤗 Simulate for reinforcement learning, continue with reinforcement learning how-tos.