Validate sampled instances for post-load stability

[cgokmen]

Example from external user:

import os
import yaml
import omnigibson as og
from omnigibson.macros import gm
gm.USE_GPU_DYNAMICS = True
from omnigibson.object_states import *
from omnigibson.systems import get_system

# Load the config
config_filename = os.path.join(og.example_config_path, "tiago_primitives.yaml")
config = yaml.load(open(config_filename, "r"), Loader=yaml.FullLoader)

config["scene"]["scene_model"] = "Beechwood_0_int"
config["scene"]["load_task_relevant_only"] = True
config["scene"]["not_load_object_categories"] = ["ceilings"]
config["task"] = {
    "type": "BehaviorTask",
    "activity_name": "boil_water_in_the_microwave",
    "activity_definition_id": 0,
    "activity_instance_id": 0,
    "predefined_problem": None,
    "online_object_sampling": False,
}

# Load the environment
env = og.Environment(configs=config)
scene = env.scene
robot = env.robots[0]

for _ in range(20):
    og.sim.step()

mug = scene.object_registry("name", "mug_188")
print(mug.get_orientation())

The mug tips over as soon as simulation starts.

Why is this happening? Can we reject these samples / implement stability checks?