World not rendering / visualizing on GPU?

Omniverse Isaac Sim
1

When I run the code below on the CPU, things work as expected. However, when I run it on GPU, the robots do not seem to move at all! However, .get_joint_state() shows they moved, which means the app is not rendering their joint states.

How can I render robots’ movements when running the simulation on GPU?

The screencast shows what I tried to explain in the post. I attach robots’ USD (in the .zip file because the forum doesn’t support .USD files upload) and a minimal reproducible example.

As a side note, based on this answer, I tried adding world.get_physics_context().enable_fabric(True) but that didn’t make a difference.

robot.zip (1.2 MB)

from isaacsim import SimulationApp

simulation_app = SimulationApp({"headless":False})

from omni.isaac.core.world import World
from omni.isaac.core.utils.stage import add_reference_to_stage
from omni.isaac.core.prims import XFormPrim
from omni.isaac.core.articulations import ArticulationView
from omni.isaac.core.utils.types import ArticulationAction
import torch
import numpy as np

use_gpu = True
if use_gpu:
    device='cuda:0'
    backend='torch'
else:
    device='cpu'
    backend='numpy'

batch_size = 4

world=World(device=device, backend=backend)
if 'cuda' in device:
    world.get_physics_context().enable_gpu_dynamics(True)
world.scene.add_default_ground_plane()

# add the robots
for i in range(batch_size):
    robot_str = "/World/Robot_" + str(i + 1)
    add_reference_to_stage(usd_path='robot.usd', prim_path=robot_str)
    position = np.array([i,0,0])
    if 'cuda' in backend:
        position = torch.tensor(position).to(device)
    world.scene.add(XFormPrim(prim_path=robot_str, name=robot_str, position=position))

world.initialize_physics()

robot_range = "/World/Robot_[1-9]|[1-9][0-9]{1,3}|9000"
robots_view = ArticulationView(prim_paths_expr=robot_range, name="robots_view")

robots_view.initialize()
world.scene.add(robots_view)
world.play()

q = torch.rand((batch_size, 23)).to(device)

if 'numpy' in backend:
    q = q.cpu().numpy()

print("target q: ", q)

for i in range(240):
    robots_view.apply_action(ArticulationAction(joint_positions=q))
    world.step(render=True)
    # same thing
    #robots_view.set_joint_positions(q)
    states = robots_view.get_joints_state()
    if i % 24 == 0:
        print(states.positions)

while True:
    world.step(render=True)
4 Likes
2

You may want to try world.set_gpu_dynamics_enabled(enabled) instead, as explained in our speed-ups doc.

4

I just pip installed the latest available Isaac sim (4.2.0.2) and that function is not supported

AttributeError: 'World' object has no attribute 'enable_gpu_dynamics'

world.get_physics_context().enable_gpu_dynamics(True) is supported, but result is the same

FWIW I printed out:

    print(world.get_physics_context().device)
    print(world.get_physics_context().is_gpu_dynamics_enabled())
    print(world.get_physics_context().use_gpu_pipeline)
    print(world.get_physics_context().use_fabric)
    print(world.get_physics_context().use_gpu_sim)

and got

cuda:0
True
True
False
True

even if I’d set world.get_physics_context().enable_fabric(True)