8000 GitHub - irmakguzey/see-to-touch: Code base for See to Touch project: https://see-to-touch.github.io/
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

irmakguzey/see-to-touch

BranchesTags

Repository files navigation

See to Touch: Learning Tactile Dexterity through Visual Incentives

[Paper] [Project Website] [Data]

Authors: Irmak Guzey, Yinlong Dai, Ben Evans, Soumith Chintala and Lerrel Pinto, New York University and Meta AI

This repository includes the official implementation of TAVI. It includes online imitation learning (IL) algorithm used with implementations of offline IL, representation learning algorithms, RL agents, exploration and reward calculation modules. Our setup tackles 6 different dexterous manipulation tasks shown above and uses an Allegro hand with XELA sensors integration and Kinova arm as the hardware.

Demonstrations are collected through the Holo-Dex pipeline and they are public in this Google Drive link.

Getting started

The following assumes our current working directory is the root folder of this project repository; tested on Ubuntu 20.04 LTS (amd64).

Setting up the project environments

NOTE: This codebase is dependant on a different version of Holo-Dex (holobot) framework which is not yet publicized and few of the packages such as AllegroKDL should be separately installed from Holo-Dex repository.

  • Install the project environment: 
    conda env create --file=conda_env.yml
    This will create a conda environment with the name see_to_touch.
  • Activate the environment: 
    conda activate see_to_touch
  • Install the see_to_touch package by using setup.py. 
    pip install -e .
    This command should be done inside the conda environment. You can test if the project package has been installed correctly by running import see_to_touch from a python shell.
  • Install the dexterous_env package by using setup.py. This package involves the created robot environments. 
    cd envs
pip install -e .
    dexterous_env directory is located under envs directory from the root of this repo.
  • To enable logging, log in with a wandb account: 
    wandb login

Downloading the demonstration dataset

As mentioned the task demonstrations are public in this Google Drive link. After compressing the .zip file 6 different demonstrations for each task can be reached through:

<download-location>/see_to_touch/tasks/<bowl_unstacking|eraser_turning|plier_picking|sponge_flipping|peg_insertion|mint_opening>

Depending on the downloaded location, data_path variable in the config of task files should be updated. (Currently for each task it is set to /see_to_touch/tasks/<task-name> as can be seen here).

Reproducing experiments

First you should activate the conda environment by running: conda activate see_to_touch.

Preprocessing

Before starting the training, each task data should be subsampled with respect to the total distance changed in the fingertips of the robot hand and the end effector position of the arm.

This can be done by running python preprocess.py. Necessary configurations should be made to the see_to_touch/configs/preprocess.yaml , such as the distances used for subsampling and which modalities (image, tactile, robot and etc.) to use in the system.

SSL Pretraining

TAVI uses a new self-supervised representation learning algorithm that utilizes the temporal change and the object / robot configuration from the demonstrations. It does this by combining an InfoNCE loss with positive pairs from nearby observations and the prediction loss of the state change between these observation.

These visual representations can be train by running:

python train.py encoder=temporal_ssl learner=temporal_ssl dataset=temporal_joint_diff_dataset data_dir=<task-data-path>

The model weights are saved under the hydra experiment created and should be used on the downstream online training.

We include the trainings of different visual representations mentioned in the paper as well. These can be trained by changing the encoder, learner and dataset variables of the see_to_touch/configs/train.yaml config file.

Downloading the tactile encoder weights

TAVI uses tactile encoders trained in T-Dex from play data. This encoder weights are public here. After downloading this folder, tactile_encoder folder path should be set as tactile_out_dir in see_to_touch/configs/train_online.yaml config file.

Online Imitation Learning

TAVI uses DrQv2 as the main reinforcement learning (RL) agent and the rewards are calculated using the optimal-transport calculation between the last frame of the demonstration and the last 10 frames of the robot trajectory.

We have different implementations of each of the module in our setup.

After finishing the previous steps and setting the image_out_dir variable to the hydra experiment folder path from SSL pretraining for each task (for example look here), you can start the online training as:

python train_online.py agent=tavi rl_learner=<drqv2|drq> base_policy=<openloop|vinn|vinn_openloop> explorer=ou_noise rewarder=<sinkhorn_cosine|sinkhorn_euclidean|cosine|euclidean> task=<task-name>

First value of each variable is as how it is in the publicized version of our implementation. We included the possible experimentations that can be done in this command as well.

Also, episode_frame_matches and expert_frame_matches can be modified to experiment further with the frames to include in reward calculation.

Citation

If you use this repo in your research, please consider citing the paper as follows:

@misc{guzey2023touch,
    title={See to Touch: Learning Tactile Dexterity through Visual Incentives}, 
    author={Irmak Guzey and Yinlong Dai and Ben Evans and Soumith Chintala and Lerrel Pinto},
    year={2023},
    eprint={2309.12300},
    archivePrefix={arXiv},
    primaryClass={cs.RO}
}

About

Code base for See to Touch project: https://see-to-touch.github.io/

Resources

Readme
Activity

Stars

48 stars

Watchers

1 watching

Forks

6 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages

0