Fine-Tuning with Lossy Affordance Planner (FLAP)

This repo contains the code for:

Generalization with Lossy Affordances: Leveraging Broad Offline Data for Learning Visuomotor Tasks

Kuan Fang, Patrick Yin, Ashvin Nair, Homer Walke, Gengchen Yan, Sergey Levine.

Conference on Robot Learning (CoRL), 2022. Oral Presentation

[Project Page]

BibTex:

@article{fang2022flap,
      title={Generalization with Lossy Affordances: Leveraging Broad Offline Data for Learning Visuomotor Tasks}, 
      author={Kuan Fang and Patrick Yin and Ashvin Nair and Homer Walke and Gengchen Yan and Sergey Levine},
      journal={Conference on Robot Learning (CoRL)}, 
      year={2022},
}

Installation

Create Conda Env

Install and use the included anaconda environment.

$ conda env create -f flap.yml
$ source activate flap
(flap) $

Dependencies

Download the dependency repos.

• bullet-manipulation (contains environments): git clone https://github.com/patrickhaoy/bullet-manipulation.git
• multiworld (contains environments): git clone https://github.com/vitchyr/multiworld
• rllab (contains visualization code): git clone https://github.com/rll/rllab

Add paths.

export PYTHONPATH=$PYTHONPATH:/path/to/multiworld
export PYTHONPATH=$PYTHONPATH:/path/to/doodad
export PYTHONPATH=$PYTHONPATH:/path/to/bullet-manipulation
export PYTHONPATH=$PYTHONPATH:/path/to/bullet-manipulation/bullet-manipulation/roboverse/envs/assets/bullet-objects
export PYTHONPATH=$PYTHONPATH:/path/to/railrl-private

Offline Dataset and Goals

Below we assume the data is stored at DATA_PATH.

Download the simulation data and goals from here. Alternatively, you can recollect a new dataset by running in the root directory of bullet-manipulation:

python shapenet_scripts/4dof_rotate_td_pnp_push_color_angle_vary_demo_collector_parallel_mixedtask.py --save_path DATA_PATH/ --name drawer_pnp_push_vary_color_angle_mixed_tasks --num_threads 4

and resample new goals by running:

python shapenet_scripts/presample_goal_with_plan.py --output_dir DATA_PATH/drawer_pnp_push_vary_color_angle_mixed_tasks/ --downsample --test_env_seeds 0 1 2 --timeout_k_steps_after_done 5 --mix_timeout_k

Experiments

Below we assume the trained models are stored at DATA_CKPT.

Offline Pre-Training

To pretrain a VQ-VAE on the simulation dataset, run:

python experiments/train_eval_vqvae.py --data_dir DATA_PATH/drawer_pnp_push_vary_color_angle_mixed_tasks --root_dir DATA_CKPT/flap/vqvae

To visualize loss curves and reconstructions of images with VQ-VAE, open the tensorboard log file with tensorboard --logdir DATA_CKPT/flap/vqvae.

To train offline RL, run:

python experiments/train_eval_flap.py 
--data_dir DATA_PATH --local --gpu --save_pretrained 
--name flap/offline_rl

To train affordance, run:

python experiments/train_eval_affordance.py 
--data_dir DATA_PATH --local --gpu --save_pretrained 
--name flap/affordance
--pretrained_rl_path DATA_CKPT/flap/offline_rl/run0/id0/itr_-1.pt

Online Fine-Tuning

To conduct online fine-tuning with FLAP, run:

python experiments/train_eval_flap.py
--data_dir DATA_PATH --local --gpu --save_pretrained
--name flap/flap
--pretrained_rl_path DATA_CKPT/flap/affordance/run0/id0/itr_-1.pt
--arg_binding algo_kwargs.start_epoch=0
--arg_binding algo_kwargs.num_epochs=150
--arg_binding reward_kwargs.obs_type='image' 
--arg_binding reward_kwargs.reward_type='cosine'
--arg_binding eval_contextual_env_kwargs.fraction_planning=0.0
--arg_binding finetune_with_obs_encoder=True
--arg_binding eval_seeds=0

Visualization

During training, the results will be saved to a file called under

LOCAL_LOG_DIR/<exp_prefix>/<foldername>

• LOCAL_LOG_DIR is the directory set by railrl.launchers.config.LOCAL_LOG_DIR
• <exp_prefix> is given either to setup_logger.
• <foldername> is auto-generated and based off of exp_prefix.
• inside this folder, you should see a file called params.pkl. To visualize a policy, run

You can visualize the results by running jupyter notebook, opening flap_reproduce.ipynb, and setting dirs = [LOCAL_LOG_DIR/<exp_prefix>/<foldername>].