Repository providing the source code for the paper
DITTO: Demonstration Imitation by Trajectory Transformation
Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox and Abhinav Valada
We recommend to use conda.
# conda create -n DITTO python=3.8 # for a ROS-compatible environment
conda create -n DITTO python=3.10 # future proof
conda activate DITTO
Install torch as your liking/your CUDA version e.g.
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt
conda install -c conda-forge libstdcxx-ng
pip install -e .
Frist clone the fork of the repo
git clone [email protected]:SuperN1ck/flowcontrol.git
follow the instructions in the readme, i.e.
pip install -e .
and installing RAFT as described here
Download the all files from here and extract them
tar -xzvf demonstration_data.tar.gz
at a location to your liking. As defined in config.py, per default DITTO assumes demonstrations located under the project root under demonstrations/.
If correctly extrated you should see a file structure like
demonstrations/
recordings_24_01_04/
coke_tray/
000/
cnos_redetected_masks/ # We used CNOS to redect all masks given the initial frame.
object_seg_000.png
...
confidence/ # Confidence arrays
00000.npy
...
depth/ # Depth arrays
00000.npy
...
hands23/ # Output of hands23, see their repo (link below)
left_hand/
right_hand/
00000.png
...
rgb/ # RGB arrays
00000.png
...
sam_refined/
# Refined masks using our procedure described in the appendix.
# Provides a container segmentation or object segmentation for the timestep in the name.
# Not for every timestep.
[container_seg_000.png]
[object_seg_000.png]
...
config.yaml # Zed2 Camera Configuration
images.np.npz # All images compressed into a single archive for faster loading
intrinsics.yaml # The instrinsics of the used camera
time_steps.yaml # See down low for a thorough explanation
zed2.svo # Raw Zed2 recordings
001/
...
...
recordings_25_01_05/
...
For an overview of the Hands23 format see their Github Page.
All scripts are powered by tyro, an awesome CLI wrapper, thus, you can simply append --help and will get an overview of all parameters.
The scripts have been run in a while and could be potentially be broken/out of sync with their respective. We provide pre-processed masks in our data already. If you have any issues and/or suggestions, happy for any issue/PR!
To extract labels use
python scripts/process_with_hands23.py
To then automatically annotate timesteps based on the hand state use
python scripts/auto_annotate_timesteps.py
We provide two methods to improve object and goal segmentation masks. First, we use CNOS to re-detect objects
python scripts/redetect_with_cnos.py
And second, we use SAM and spatial distances
python scripts/refine_with_sam.py
We found the latter to perform consistently better.
We provide a notebook under notebooks/DITTO_example_usage.ipynb that showcases how to use DITTO.
To run the same evaluations as in the paper use the following commands.
python scripts/eval/track_correspondences_within_episode.py
python scripts/eval/track_correspondences_between_episode.py
python scripts/eval/evaluate_relative_poses_between_trajectories.py
The previous mentioned scripts will produce result files. For investigating the results we provide two notebooks under
notebooks/
eval/
correspondences.ipynb
poses.ipynb
We also provide a notebook to render rotating 3D scenes under notebooks/show_single_trajectory.ipynb.
For each demonstration there is a time_steps.yaml-file. It contains manual adjustements if needed.
# Automatically extracted from scripts/auto_annotate_timesteps.py
- t_start: 13
- t_stop: 44
# Use the CNOS-based re-dection mask over Hands23 ones (CNOS is very unlikely to be better due to occlusions)
- cnos_over_hands23: [17, ...]
# In case none of our methods produced a good segmentation mask, we will shift a timestep back/forward
- redirect_to_other_step: {21: 22, ...}
# Overwrites the start of the episode
- t_h23_object_manual: 11
# Overwrites the stop of the episode
- t_stop_manual: 35
If you find this repository useful, please consider citing the paper as follows ✍🏼:
@inproceedings{heppert2024ditto,
title={DITTO: Demonstration Imitation by Trajectory Transformation},
author={Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox, Abhinav Valada},
booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2024},
organization={IEEE}
}
Also checkout our utility library casino.
Please note that different from the paper, we split the tennisball_cup into two seperate tasks.
Original, for the numbers in the paper, they were all merged under one
tennisball_cup/
000/
001/
002/
003/
004/
In this repository, as defined in config/valid_episodes.yaml, we assume a split into
tennisball_cup_laying/
000/
001/
002/
tennisball_cup_upright/
003/
004/
Vilja Lott also developed an updated metric for calculating the relative pose error. We will release the results soon.
This work was partially funded by the Carl Zeiss Foundation with the ReScaLe project.