README.md

Imitation from Observation with Bootstrapped Contrastive Learning

Imitation from observation algorithm to train agents to perform tasks using only a limited number of pixel-based expert observations and based on a behavioral learning principle.

An encoder that takes videos of agent trajectories and embeds them in a "behavioral space" is trained using contrastive learning (enforcing successful trajectories to lie close together). We use this to encode N expert videos in a region of the behavioral space in blue. The reward function corresponds to the distance of the agent's trajectory to the set of expert trajectories. As the agent progresses, its current trajectories are incorporated as "negative" examples into the contrastive learning in red.

Demonstration videos. The expert on the left and the IfO agent on the right.

Installation

conda env create -f env.yml
conda activate ifobl

Training (example with reacher hard task)

• Train expert

python train.py task=reacher_hard exp_group=reacher_hard exp_id=1

Watch training on tensorboard

tensorboard --logdir exp_local

• Generate 5000 expert videos

export PYTHONPATH="${PYTHONPATH}:`pwd`" && python scripts/generate_dmc_video.py --env reacher_hard2 --episode_len 60

Use --num-train and --num-valid flags to change respectively the number of training and validation videos to generate.

• Pretrain image and video encoders

python train_cmc.py task=reacher_hard2 exp_id=1

Watch training on tensorboard

tensorboard --logdir cmc_exp_local

• Train agent

python train_rlv2.py task=reacher_hard2

Watch training on tensorboard

tensorboard --logdir rlv2_exp_local

Evaluation videos are generated in rlv2_exp_local/reacher_hard/<exp_id>/train_video directory.

Additional information

• cmc_model.py: contains the models, neural networks and losses used to train the trajectory encoder
• drqv2.py: contains the implementations of the policy and q-value functions used to train the agents and experts
• rl_model.py: contains the implementations of the policy and q-value functions used to train the state-based agents
• final_run: contains the scripts to train the experts and agents for other tasks such as Walker run, Hopper stand and finger turn
• dmc.py: contains environment creation functions and environment wrappers
• scripts/generate_dmc_video.py: shows how to use trained agents in test time

Acknowledgements

• We reuse Denis Yarats's DrQv2 code to train our RL agents