Inconsistent depth information

[pangxincheng]

Habitat-Sim version

v0.2.2

Docs and Tutorials

Did you read the docs? https://aihabitat.org/docs/habitat-sim/ Yes Did you check out the tutorials? https://aihabitat.org/tutorial/2020/ Yes Perhaps your question is answered there. If not, carry on!

❓ Questions and Help

I recently did an experiment on Habitat sim. The agent first turned 30 ° to the left, then turned 30 ° to the right (back to the original state). However, the depth is inconsistent after such operation.

Here's my code.

import random
import numpy as np
import habitat_sim

def make_cfg(settings):
    sim_cfg = habitat_sim.SimulatorConfiguration()
    sim_cfg.gpu_device_id = settings["gpu_device_id"]
    sim_cfg.scene_id = settings["scene"]
    sim_cfg.enable_physics = settings["enable_physics"]
    sim_cfg.scene_dataset_config_file = settings["scene_dataset"]

    sensor_specs = []

    if settings["rgb_sensor"]:
        rgb_sensor_spec = habitat_sim.CameraSensorSpec()
        rgb_sensor_spec.uuid = "rgb_sensor"
        rgb_sensor_spec.hfov = settings["hfov"]
        rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR
        rgb_sensor_spec.resolution = [settings["height"], settings["width"]]
        rgb_sensor_spec.position = [0.0, settings["sensor_height"], 0.0]
        rgb_sensor_spec.sensor_subtype = habitat_sim.SensorSubType.PINHOLE
        sensor_specs.append(rgb_sensor_spec)

    if settings["depth_sensor"]:
        depth_sensor_spec = habitat_sim.CameraSensorSpec()
        depth_sensor_spec.uuid = "depth_sensor"
        depth_sensor_spec.hfov = settings["hfov"]
        depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH
        depth_sensor_spec.resolution = [settings["height"], settings["width"]]
        depth_sensor_spec.position = [0.0, settings["sensor_height"], 0.0]
        depth_sensor_spec.sensor_subtype = habitat_sim.SensorSubType.PINHOLE
        depth_sensor_spec.noise_model = "None"
        sensor_specs.append(depth_sensor_spec)

    if settings["semantic_sensor"]:
        semantic_sensor_spec = habitat_sim.CameraSensorSpec()
        semantic_sensor_spec.uuid = "semantic_sensor"
        semantic_sensor_spec.hfov = settings["hfov"]
        semantic_sensor_spec.sensor_type = habitat_sim.SensorType.SEMANTIC
        semantic_sensor_spec.resolution = [settings["height"], settings["width"]]
        semantic_sensor_spec.position = [0.0, settings["sensor_height"], 0.0]
        semantic_sensor_spec.sensor_subtype = habitat_sim.SensorSubType.PINHOLE
        sensor_specs.append(semantic_sensor_spec)

    agent_cfg = habitat_sim.agent.AgentConfiguration()
    agent_cfg.sensor_specifications = sensor_specs
    agent_cfg.action_space = {
        "move_forward": habitat_sim.agent.ActionSpec(
            "move_forward", habitat_sim.agent.ActuationSpec(amount=0.25)
        ),
        "turn_left": habitat_sim.agent.ActionSpec(
            "turn_left", habitat_sim.agent.ActuationSpec(amount=30.0)
        ),
        "turn_right": habitat_sim.agent.ActionSpec(
            "turn_right", habitat_sim.agent.ActuationSpec(amount=30.0)
        ),
    }

    return habitat_sim.Configuration(sim_cfg, [agent_cfg])

def main():
    sim_settings = {
        "width": 256,
        "height": 256,
        "hfov": 90,
        "scene": "data/scene_datasets/mp3d/17DRP5sb8fy/17DRP5sb8fy.glb",
        "scene_dataset": "data/scene_datasets/mp3d/mp3d.scene_dataset_config.json",
        "default_agent": 0,
        "sensor_height": 1.5,
        "rgb_sensor": False,
        "depth_sensor": True,
        "semantic_sensor": False,
        "seed": 1,
        "enable_physics": False,
        "gpu_device_id": 0
    }

    cfg = make_cfg(sim_settings)
    sim = habitat_sim.Simulator(cfg)

    random.seed(sim_settings["seed"])
    sim.seed(sim_settings["seed"])

    agent = sim.initialize_agent(sim_settings["default_agent"])
    agent_state = habitat_sim.AgentState()
    agent_state.position = np.array([-0.6, 0.0, 0.0])
    agent.set_state(agent_state)

    observations = sim.reset(0)
    depth_raw = observations["depth_sensor"]

    actions_seq = ["turn_left", "turn_right"]

    for action in actions_seq:
        observations = sim.step(action)
        depth_new = observations["depth_sensor"]

    if not (depth_raw == depth_new).all():
        # Issue: how to ensure consistent observation
        not_eq_idx = np.argwhere(depth_new != depth_raw)
        print(
            "len(not_eq_idx) =", 
            len(not_eq_idx), 
            "ratio =", 
            f"{100 * len(not_eq_idx) / (sim_settings['height'] * sim_settings['width'])}%")

if __name__ == "__main__":
    main()

The outputs are as follows:

len(not_eq_idx) = 2139 ratio = 3.26385498046875%

In addition, I printed inconsistent depth information, and the difference between them is very small (similar to the following). I think they may be caused by calculation errors. Can such errors be completely avoided?

depth_new 0.9272244 depth_raw 0.92721933 difference 5.066395e-06