REPPO Experiment Plan

Proof of Concept Success

✅ Working Implementation: https://wandb.ai/dominik_roth/reppo_dev_test?nw=nwuserdominik_roth

Experiments To Run

1. Reproduce Paper Results

✅ Brax Suite: 5 tasks - COMPLETED

ant, cheetah, humanoid, walker, hopper
Results: https://wandb.ai/dominik_roth/reppo_brax_production

🔄 DMC Suite (mujoco_playground): 24 tasks - IN PROGRESS

AcrobotSwingup, CartpoleBalance, CheetahRun, FingerSpin, HumanoidRun, WalkerRun, etc.
Dev test completed successfully: https://wandb.ai/dominik_roth/reppo_dmc_dev
Production jobs: 72 jobs (24 tasks × 3 seeds) submitted, IDs 3382257-3382330

ManiSkill Suite: 8 tasks (need wrapper first)

PickSingleYCB-v1, PegInsertionSide-v1, UnitreeG1TransportBox-v1, etc.

Settings: 50M steps, 1024 envs, 3 seeds each, paper hyperparameters

Scripts Available

DMC Experiments

# Submit all 24 DMC tasks with 3 seeds each
python submit_dmc_experiments.py --seeds 3

# Submit specific tasks
python submit_dmc_experiments.py --tasks CartpoleBalance CheetahRun --seeds 3

Brax Experiments (Already completed)

./submit_job.sh brax ant mjx_dmc_medium_data
./submit_job.sh brax cheetah mjx_dmc_medium_data

1.2 KiB Raw Blame History Unescape Escape