HouseExpo Dataset & PseudoSLAM Simulator (IROS2020)

by Tingguang Li at The Chinese University of Hong Kong. The paper and video can be found at Paper, Video.

If you think our work is useful, please consider citing use with

@article{"li2019houseexpo",
  title={HouseExpo: A Large-scale 2D Indoor Layout Dataset for Learning-based Algorithms on Mobile Robots},
  author={Tingguang, Li and Danny, Ho and Chenming, Li and Delong, Zhu and Chaoqun, Wang and Max Q.-H. Meng},
  journal={arXiv preprint arXiv:1903.09845},
  year={2019}
}

Overview

HouseExpo is a large-scale dataset of indoor layout built on SUNCG dataset. The dataset contains 35,126 2D floor plans with 252,550 rooms in total, together with the category label of each room. Check out all floor plans as .png images at: https://drive.google.com/file/d/1gEmTdgZD1pIa8UtaLXz8vm_301zL4L1J/view?usp=sharing.

PseudoSLAM is a high-speed OpenAI Gym-compatible simulation platform that simulates SLAM and the navigation process in an unknown 2D environment. It reads data from HouseExpo, creates the corresponding 2D environment and generates a mobile robot to carry on different tasks in this environment.

Prerequisite

The code has been tested under

• python 3.6
• tensorflow 1.15
• Ubuntu 16.04

Getting Started

• Clone the repo and cd into it:

  git clone https://github.com/TeaganLi/HouseExpo.git
  cd HouseExpo
  

• Install pseudoSLAM package

  pip install -e .
  

• Uncompress HouseExpo data:

  cd HouseExpo
  tar -xvzf json.tar.gz
  

• There are two demos for you. You can try the exploration demo using keyboard by running

  python pseudoslam/envs/keyboard_exploration.py
  

  You will see an exploration (partially known/unknown) environment (left figure). Activate the terminal and then you can control the robot by pressing w (move forward), a (rotate clockwise) and d (rotate anti-clockwise).

• You can try the navigation demo using keyboard by running

  python pseudoslam/envs/keyboard_navigation.py
  

  You will see a navigation (fully known) environment. Activate the terminal and you can save the traversed trajectory by pressing s.

• To train your model, you can check the observation and adjust your configuration file by running the following code. You will see a global map and a local observation.

  python pseudoslam/envs/robot_exploration_v0.py
  

Training Models

• One Gym-compatible environment of robot exploration has been implemented. To train a model using OpenAI baselines, first install OpenAI Gym and Baselines. Then add our environment in Gym by registering in /gym/envs/__init__.py:

  register(
      id='RobotExploration-v0',
      entry_point='pseudoslam.envs.robot_exploration_v0:RobotExplorationT0',
      max_episode_steps=200,
  )
  

• Start to train the model using baselines, for example, run

  python -m baselines.run --alg=ppo2 --env=RobotExploration-v0 --network=cnn --num_timesteps=2e7 --save_path=~/models/exploration_20M_ppo2 --save_interval=10 --log_path=~/logs/exploration --save_video_interval=10000
  

• After the training is finished, check your trained model by running

  python -m baselines.run --alg=ppo2 --env=RobotExploration-v0 --network=cnn --num_timesteps=0 --load_path=~/models/exploration_20M_ppo2 --play
  

Please refer to https://github.com/openai/baselines for more detailed introduction on how to use baselines, like launching multiple environments.

Details of HouseExpo Dataset

Data Format

The floor plans are stored in ./HouseExpo/json/ in the form of .json files. The data format is as follows

• id (string): the unique house ID number.
• room_num (int): the number of rooms of this house.
• bbox (dict): bounding box of the whole house
  • "min": (x1, y1)
  • "max": (x2, y2)
• verts (list): each element (x, y) represents a vertex location (in meter).
• room_category (dict): the room categories and its bounding box, for example
  • "kitchen": (x1, y1, x2, y2), bounding box of each kitchen.

Map Visualization

First randomly sample a subset of maps by running

python pseudoslam/viz/map_id_set_generator.py --num 100

Then visualize the sampled maps as images

python pseudoslam/viz/vis_maps.py PATH_TO_MAPID_FILE

Room Category

HouseExpo inherits room category labels from SUNCG dataset and also provides a flexible way to define your own room type labels (as defined in pseudoslam/envs/simulator/jsonreader.py). Some samples are (different colors represent different room categories)

Statistics

There are 35,126 2D floor plans with 252,550 rooms, with mean of 7.14 and median of 7.0 rooms per house. The distribution of rooms is

Feel free to download all floor plans as .png images at: https://drive.google.com/file/d/1gEmTdgZD1pIa8UtaLXz8vm_301zL4L1J/view?usp=sharing

Usage of PseudoSLAM

Simulator Parameters

The parameters of the simulator are specified in ./pseudoslam/envs/config.yaml including

• json_dir: path to json files
• map_id_set: path to map_id_set. PseudoSLAM supports to utilize a subset of the dataset to train a model. map_id_set is a .txt file recording all map ids you want to train/test on.
• meter2pixel: the ratio from meter to pixel, i.e. 1 meter in real world corresponds to n pixels in simulator. Note that a larger value will present a more precise simulation, but with a low simulation speed.
• mode: 0 for exploration (unknown environment), 1 for navigation (fully known environment)
• obstacle
  • number: number of obstacles. The obstacle is dynamically generated when initializing the map.
  • size [max, min]: size of obstacle (in meter).
• robotRadius: robot radius (in meter) for collision checking.
• stepLength:
  • linear: robot linear movement in each step (in meter).
  • angular: robot angular movement in each step (in degree).
• startPose [x, y, theta]: robot start pose (in meter) in world coordinate with (0,0) at the center of the house.
• resetRandomPose: flag of whether randomly set robot init pose when reset, if 0, robot reset to startPose.
• laser:
  • range: laser range in meter.
  • fov: laser field of view in degree.
  • resolution: laser rotation resolution in degree.
  • noiseSigma: sigma of Gaussian distributino of laser noise.
• slamError:
  • linear: slam linear registration error in pixel.
  • angular: slam rotational registration error in degree.
• stateSize [x, y]: state size in meter

Benchmark Reproduction

The simulation configurations and the training/testing map ids for obstacle avoidance and autonomous exploration are located at ./experiments for reproduction purpose.

Dynamic Obstacles

We collected a database for dynamic obstacles (moving humans) from the real world and are incorporating this part into our simulator. A demo is like this. Please stay tuned for this part!