[Bug]: scene.step() gets stuck indefinitely

[spen2005]

Bug Description

While simulating a robot hand interacting with objects, the call to scene.step() sometimes gets stuck indefinitely. When this happens, the process becomes unresponsive and cannot be terminated even with Ctrl-C.

This bug can be reproduced using the provided script in the earlier version, but not in the latest version. However, the issue still occurs in the latest version—just under different simulation conditions, especially when the robot hand has complex contact interactions with objects.

Steps to Reproduce

Please download the action sequence and object files in bug_report2.zip

bug_report2.zip

import genesis as gs
import numpy as np
import torch

gs.init(seed=42, backend=gs.gpu)

scene = gs.Scene(
            show_viewer = True,
            sim_options=gs.options.SimOptions(
                dt       = 1 / 60,
                substeps = 4,
            ),
            viewer_options = gs.options.ViewerOptions(
                camera_pos =  (0, 0.3, 1.6),
                camera_lookat = (0, 0.47, 0.7),
            ),
            rigid_options = gs.options.RigidOptions(
                dt = 1 / 60,
                enable_self_collision = False,
                use_contact_island=False,  # Note: for multiple-contact
                constraint_solver=gs.constraint_solver.Newton,
                tolerance=1e-8,
                iterations=1000,
                ls_iterations=500,
                max_collision_pairs=1000,
            ),
            vis_options = gs.options.VisOptions(
                env_separate_rigid = True,
            )
        )

plane = scene.add_entity(
            gs.morphs.Plane(),
        )

object = scene.add_entity(
            morph=gs.morphs.URDF(
                file="./assets/C12001/C12001.urdf",
                pos=(2, 0, 0), # Initial pos set during reset
                fixed=False,
                convexify=True,
                coacd_options=gs.options.CoacdOptions(
                    threshold=0.3,
                ),
            ),
        )

robot = scene.add_entity(
            gs.morphs.URDF(
                file="./assets/g1_shadow/g1_29dof_shadow.urdf",
                pos=(0, 0, 0.75),
                quat = (0.7071, 0, 0, 0.7071),
                fixed=True,
                merge_fixed_links=True,
                convexify=True,
                decompose_robot_error_threshold=float("inf"),
            ),
        )

robot.set_friction(0.5)

table = scene.add_entity(
            gs.morphs.Box(
                pos = (0, 0.9, 0.02),
                size = (1.5, 1.5, 0.04),
                fixed = True,
            )
        )

scene.build(n_envs = 1, env_spacing=(0,0), center_envs_at_origin = False)

robot_kp = np.load("robot_kp.npy")
robot_kv = np.load("robot_kv.npy")
robot.set_dofs_kp(torch.tensor(robot_kp, device=gs.device))
robot.set_dofs_kv(torch.tensor(robot_kv, device=gs.device))

robot_init_action = np.load("init_action.npy")
robot.set_dofs_position(torch.tensor(robot_init_action, device=gs.device))
print("init action", torch.tensor(robot_init_action, device=gs.device))

init_table_pos = np.load("init_table_pos.npy")
table.set_pos(torch.tensor(init_table_pos, device=gs.device))

init_bowl_pos = np.load("init_bowl_pos.npy")
object.set_pos(torch.tensor(init_bowl_pos, device=gs.device))

init_bowl_quat = np.load("init_bowl_quat.npy")
object.set_quat(torch.tensor(init_bowl_quat, device=gs.device))

for i in range(16):
    action = np.load(f"action{i+1}.npy")
    robot.control_dofs_position(torch.tensor(action, device=gs.device))
    scene.step()

Expected Behavior

scene.step() should proceed normally without freezing.

Screenshots/Videos

No response

Relevant log output

The process appears to hang at the following line:

File "collider_decomp.py", line 269 in detection

Environment

OS: Ubuntu 24.04.1 LTS GPU: GeForce RTX 4090 CPU: AMD Ryzen 9 7950X 16-Core Processor GPU-driver: 560.28.03 CUDA / CUDA-toolkit version: 12.4

Release version or Commit ID

01f0858997767cb82b29bd0fcdcacab278dc40fc

Additional Context

Thanks for looking into this! Really appreciate your help.