Authors: Albert Wu(amhwu [at] stanford [dot] edu), Ruocheng Wang, Sirui Chen, Clemens Eppner, C. Karen Liu.

This repository contains the code for the paper One-Shot Transfer of Long-Horizon Extrinsic Manipulation Through Contact Retargeting

Installation

Prerequesites

Set up conda virtual environment

This repository is tested in python 3.8.

conda create -n "extrinsic-manip-release-3.8" python=3.8 ipython

Install isaacgym

Obtain the preview release and install following the instructions.

Installing the main code

cd contact_demo
pip install -e .

Additional Downloads

Object scans

Scans of the standard and short objects. We suggest extracting the scans to assets/urdf/tml/.

Pretrained pushing reinforcement learning policy

The pretrained weights for the pushing reinforcement learning policy is available here

We suggest placing the pretrained policy in assets/nn.

Precomputed initial-final state pairs for training pushing policy

The initial and final state pairs we used to train the pushing policy is available here. Alternatively, you can generate the data using the instructions below.

We suggest placing the precomputed state pairs in assets/state_pairs/.

Repository structure

The three main directories relevant to the users are: assets, contactdemo, isaacgymenvs

assets

This directory contains the object scans, pretrained policy, precomputed state pairs, relevant CAD files, and other assets. Refer to the Additional Downloads section for what should go in this directory.

contactdemo

This directory contains the non-learning code for the paper.

contactdemo/lib/data_generation contains the code for generating the initial and final states for the pushing task.

contactdemo/lib/drake/contact_retargeting contains the implementation of Section IV.B Implementing $\sigma$ of the paper using Drake.

contactdemo/scripts/demo/record_demo_obj_path.py is a simple script for recording the demo object path (hardware required).

isaacgymenvs

This is adapted from IsaacGymEnvs. It implements the reinforcement learning pipeline for the pushing task and the multi-stage policy.

The class HardwareVecTask in isaacgymenvs/tasks/base/vec_task.py has been heavily modified to implement Algorithm 1 of the paper. It is the backbone of the multi-stage framework, in addition to still supporting the single-stage training for the pushing policy.

FrankaYCBPush is the main "task" for both training the pushing policy.

FrankaYCBUngraspable is a slightly modified version of FrankaYCBPush for end-to-end training of grasping the ungraspable (Section VI.C of the paper).

Running the code

Simulation cannot capture the extrinsic manipulation dynamics. Moreover, lots of our code requires real-time state feedback from the robot and perception system. Thus we only provide instructions for running a very limited subset of our code. Please contact the authors if you want to run the pipeline on hardware.

Pushing Data Generation

To generate the initial and end states of objects for training the pushing task, run the following command:

 PYTHONPATH=. python contactdemo/lib/data_generation/general_pushing_data_generation_scanned_parallel.py --dataset_path $SOME_DIRECTORY --n_examples $N 

The dataset will be saved to $SOME_DIRECTORY.pkl, where the data is stored as:

{
  'mustard': [
    {
        'initial_state':  [...], # 7D pose (3D position, 4D quaternion)
        'final_state':  [...] # 7D pose (3D position, 4D quaternion)
    },
    ...
  'xxx': [...]
   ...
}

After the data is generated, you can visualize the data by running the following command:

 PYTHONPATH=. python contactdemo/lib/data_generation/trajectory_visualization_scanned.py --pkl_path $SOME_DIRECTORY.pkl

Visualizing the pretrained push policy

After downloading the object scans, pretrained policy, and precomputed state pairs, run

python train.py task=FrankaYCBPush test=True num_envs=$NUM_ENVS checkpoint=$PATH_TO_PRETRAINED_MODEL wandb_activate=False

The number of environments should be set to at least 10 to see all the objects. The $i$th object appears in environments $j$ where $j\%10 = i$.

Please refer to section Configuration and command line arguments in isaacgymenvs_docs/README.md for details on the command line arguments. Relevant parameters can be found in isaacgymenvs/cfg/task/FrankaYCBPush.yaml.