From Designer to Simulation

The browser designer creates geometry and connectivity. Python runs the physics. The browser does not train controllers or preview a physical rollout.

Prepare and export

1. Create an object named robot. Draw a connected body with at least one H-Act or V-Act cell.
2. Create a separate terrain object and draw Fixed ground below the body. Leave space for the robot to move.
3. Validate the world, resolve errors, and export evogym-world.json.

This introductory script uses scalar actions. It rejects Contractile cells because they need the per-axis control path.

Spawn, target, and direction markers are design annotations. They do not move the robot, define a reward, or configure the simulation. Initial position comes from the cells’ world coordinates. Keep your original export: the Python loader ignores metadata, and a Python JSON round trip drops it.

Run the exported world

From an installed source checkout with the visualization extra:

python -m examples.designer_world evogym-world.json --robot robot --output /tmp/world.gif

Use the exact object name from the designer. The script lists available names if it cannot find the robot. It loads the complete world, compiles a template, instantiates state, applies periodic actions, and saves a GIF. Static and dynamic collision data are passed to every simulation step.

This is a physics-only rollout: its recorded reward is zero and it defines no task completion condition. To evaluate a controller, define observations, a reward, and termination in a custom environment.

A checked-in fixture is available for testing the path without drawing first:

python -m examples.designer_world examples/world.json --output /tmp/world.gif

Use only the body in a benchmark

from jax_evogym import EvoWorld, WalkerV0

world = EvoWorld.from_json("evogym-world.json")
robot = world.objects["robot"]
env = WalkerV0(robot.get_structure(), connections=robot.get_connections())
obs, state = env.reset()

This uses the robot’s local structure and connectivity. WalkerV0 supplies its own terrain and placement; your exported terrain and markers do not configure that benchmark.

See examples/designer_world.py, the JSON reference, and the designer guide for details.