Skip to content

Latest commit

 

History

History
54 lines (42 loc) · 3.45 KB

README.md

File metadata and controls

54 lines (42 loc) · 3.45 KB
Raw

Scene Generation and Selection

Overview

The scene generation proceeds entirely in gazebo simulation by trying to spawn objects on a table in a collision free manner. For a scene, 5 objects are spawned and a motion plan is checked to each of their grasps (from an offline grasp dataset). If all 5 objects have at least 1 valid motion plan to some grasp then we select (save) the scene.

The objects are spawned in a randomly chosen stable pose along with a random rotation around Z axis applied. The table's surface is discretized into a grid with each grid point denoting the possible spawn location (x,y) for an object. If a grid point is not feasible, we search for others in a breath-first-search like manner.

Lastly, from the several scenes generated, we attempt to obtain a dataset of 20 scenes with the following properties:

  1. Each object must appear in 5-7 scenes
  2. There should be diversity in poses of an object
  3. Score a dataset by a diversity metric for stable pose used (we use entropy of count distribution)

Setup and Usage

Source the appropirate ROS workspace with all the dependencies (gazebo, fetch_ros, moveit etc). Multiple terminals needed for gazebo, moveit, main python script and rviz.

  • Setup the data directories (gazebo models, grasp data) into appropriate folders as described here
  • Gazebo and Moveit Launch:
    • Launch gazebo with just the Fetch robot and cafe_table_org model. Launching with just the robot model avoids errors in launch collisions with table.
    • Launch moveit using the launch file for simulation
  • In a third terminal, run python generate_scenes.py with appropriate command line args for data directory and number of rows and columns for the spawn grid. More rows/cols increases the number of possible object spawn locations and brings them closer (but also makes it hard to get all objects as reachable or increase possible collisions).

Scene Generation

  • generate_scenes.py: Main scripts to trigger the generation of scenes in gazebo. Check the command line args to give (main one is the data dir)

    • --data_dir : folder where a scene_gen/iter_TIMESTAMP/ folders will be created to store selected scenes
    • --rows and --cols : rows and cols used to discretize the table grid
    • --grasp : whether to also try execute the motion plan to grasp (instead of just planning it, useful for viz and debugging)

  • viz_scenes.py : Visualize the selected scenes in gazebo. The setup steps are same as that for generate_grasps.py but here you just visualize the scenes.

Saving Metadata for Scenes

This allows us to store the metadata for recreating the scene in real world

  1. Follow the above steps for gazebo and moveit.
  2. In a separate terminal, run python save_pose_results.py
    1. It will start subscribing to a topic published by viz_scenes.py and will save the appropriate data on callback
    2. Check the output dir used by this script and modify as cmd line arg if neeeded
  3. Lastly run python viz_scenes.py which will start visualizing the scenes. You can iterate over the desired scenes to save their metadata.

Selecting Final Scenes

  1. Generate 20 Datasets, each having 20 scenes
python3 select_datasets.py --min=5 --max=7 --iterations=20 --scenes_dir=/path/to/scene_pickle_files/
  1. Selecting Diverse dataset
python3 select_diverse.py  --scenes_dir=/path/to/scene_pickle_files/
  1. Visualizing a dataset
python3 plot_scenes.py