This codebase contains an implementation of the TCDM benchmark and PGDM learning framework from our paper, Learning Dexterous Manipulation from Exemplar Object Trajectories and Pre-Grasps.
If you found this code useful in any way, please cite our paper:
@InProceedings{dasari2022pgdm,
title={Learning Dexterous Manipulation from Exemplar Object Trajectories and Pre-Grasps},
author={Dasari, Sudeep and Gupta, Abhinav and Kumar, Vikash},
journal={arXiv preprint arXiv:2209.11221},
year={2022}
}
}
- Our code has been primarily tested on Ubuntu 20, but it should work on other versions of Linux
- We strongly recommend using Anaconda or Miniconda for setting up the Python environment.
- You will need a wandb account, as our code uses it for logging results.
Please run the following commands to install our codebase:
git clone --recurse-submodules [email protected]:facebookresearch/TCDM.git && cd TCDM
conda env create -f environment.yml && conda activate tcdm
pip install -r requirements.txt
python setup.py develop
export MUJOCO_GL=egl # NOTE: YOU PROBABLY WANT TO ADD THIS TO .bashrc
You should now be able to import our environment suite (python code below) and train policies. Happy experimenting!
>>> from tcdm import suite
>>> env = suite.load('hammer', 'use1')
The following examples will show how to train dexterous manipulation policies in TCDM environments, using our PGDM framework.
- Simply running
python train.pywill train an agent on thehammer-use1task. You may view the results (including videos of behaviors) on wandb. - To train other tasks, specify on command line:
python train.py env.name=<task name>. Checkout TASKS.md for a full list of our tasks, alongside goal visualizations. - Note that you can change many aspects of our training pipeline, simply by overriding our default config with command line arguments! We use hydra to handle this -- please read those docs for more information.
The following code snippet will run a sweep across all tasks in TCDM and report the results to wandb. Note: our code assumes you have access to a slurm or ray cluster! For alternate launchers check here
python train.py hydra/launcher=<slurm/ray> exp_name=tcdm_sweep wandb.project=tcdm env.name=headphones-pass1,elephant-pass1,eyeglasses-pass1,flute-pass1,banana-pass1,hand-inspect1,binoculars-pass1,stanfordbunny-inspect1,toruslarge-inspect1,alarmclock-see1,fryingpan-cook2,airplane-fly1,cup-drink1,scissors-use1,cup-pour1,mug-drink3,waterbottle-shake1,flashlight-on2,wineglass-toast1,piggybank-use1,wineglass-drink2,lightbulb-pass1,wineglass-drink1,mouse-use1,knife-chop1,airplane-pass1,duck-inspect1,hammer-use1,stamp-stamp1,train-play1,toothpaste-lift,watch-lift,toothbrush-lift,stapler-lift,mouse-lift,waterbottle-lift,spheremedium-lift,alarmclock-lift,flashlight-lift,duck-lift,dhand-waterbottle,dhand-alarmclock,dhand-cup,dhand-elephant,dhand-binoculars,dmanus-crackerbox,dmanus-coffeecan,spheremedium-relocate,door-open,hammer-strike -m
Similarly, you can sweep across all tasks in the TCDM-30 subset:
python train.py hydra/launcher=<slurm/ray> exp_name=tcdm_30_sweep wandb.project=tcdm env.name=headphones-pass1,elephant-pass1,eyeglasses-pass1,flute-pass1,banana-pass1,hand-inspect1,binoculars-pass1,stanfordbunny-inspect1,toruslarge-inspect1,alarmclock-see1,fryingpan-cook2,airplane-fly1,cup-drink1,scissors-use1,cup-pour1,mug-drink3,waterbottle-shake1,flashlight-on2,wineglass-toast1,piggybank-use1,wineglass-drink2,lightbulb-pass1,wineglass-drink1,mouse-use1,knife-chop1,airplane-pass1,duck-inspect1,hammer-use1,stamp-stamp1,train-play1 -m
Pre-trained policies are available here. Please refer to rollout.py for an example on how to load and use these policies.
If you're interested in contributing to this codebase, please refer to CONTRIBUTING.md for more information.
is MIT licensed, as found in LICENSE. However, it does rely on other libraries, including object_sim, which each have their respective licenses that must be followed.