Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Active Domain Randomization

Active Domain Randomization (ADR) is a new method for improved, zero-shot transfer of robotic reinforcement learning policies. Building upon traditional domain randomization, which uniformly samples the randomization space, we show that replacing this with an active search for difficult MDP instances improves generalization and robustness in the resulting policies.


Below is our code to reproduce the experiments in the paper. Please check out our Domain Randomizer repository if you're interested in a easy way to do domain randomization in parallel.

Experiments

Baseline Experiments

Pure Baseline

The most important flag here is --initial-svpg-steps=1e6, which will make sure that only randomized environments are proposed until that step limit is reached (it never will be). The environment names (randomized-env-id) handle the range of randomization - LunarLanderDefault-v0 has a single-valued range, so a randomize() call will always result in the same, default environment. Likewise, LunarLanderRandomized-v0 has the full randomization range (in one dimension).

On the command line, specify an experiment type from [lunar|pusher|ergo] to get defaults for that experiment. You can find a detailed list of command line arguments in experiments/args.py.

python -m experiments.domainrand.experiment_driver [lunar|pusher|ergo] \
    --experiment-name=unfreeze-policy  --freeze-discriminator \
    --experiment-prefix="true-baseline" --agent-name=baseline --initial-svpg-steps=1e6 \
    --continuous-svpg --freeze-svpg --seed={SEED}

Uniform Domain Randomization

python -m experiments.domainrand.experiment_driver [lunar|pusher|ergo] \
    --experiment-name=unfreeze-policy --randomized-eval-env-id="[corresponding env ID]" \
    --experiment-prefix="fulldr" --agent-name=fulldr --initial-svpg-steps=1e6 \
    --continuous-svpg --freeze-svpg --seed={SEED}

Active Domain Randomization

python -m experiments.domainrand.experiment_driver [lunar|pusher|ergo] \
    --experiment-name=unfreeze-policy --load-discriminator --randomized-eval-env-id="[corresponding env ID]" \
    --freeze-discriminator --experiment-prefix="ours-agent-scratch" --seed={SEED}

Reference

@article{mehta2019adr,
  title={Active Domain Randomization},
  author={Mehta, Bhairav and Diaz, Manfred and Golemo, Florian and Pal, Christopher and Paull, Liam},
  url={https://arxiv.org/abs/1904.04762},
  year={2019}
}

Built by @bhairavmehta95, @takeitallsource, and @fgolemo.

About

Code repository for Active Domain Randomization (CoRL 2019, https://arxiv.org/abs/1904.04762)

Resources

Releases

No releases published

Packages

No packages published