[Progress Report] Construction of RLLTE Data Hub

[yuanmingqi]

Due to the high computing power required for training, we will gradually upload data to the data hub and report the progress in this issue. We will also change the priority of training according to needs, and you can leave a message here.

[yuanmingqi]

Environment: DMControl
Completed:

1. Soft Actor-Critic (SAC)
   27 tasks reported in pytorch_sac. Two examples:
   • sac_dmc_state_humanoid_run (2 seeds, 10M steps)
   • sac_dmc_state_quadruped_walk (10 seeds, 2M steps)

Model import example:

from rllte.hub.models import DMControl

if __name__ == "__main__":
    model = DMControl().load_models(
        agent="sac",
        env_id="humanoid_run",
        seed=1,
        device="cuda"
    )
    print(model)

[yuanmingqi]

Environment: Envpool Atari Games synchronous mode
Completed:

1. Proximal Policy Optimization (PPO)
   57 Atari games reported in Agent57: Outperforming the Atari Human Benchmark. Two examples:
   • ppo_atari_Breakout-v5 (10 seeds, 10M steps)
   • ppo_atari_Pong_v5 (10 seeds, 10M steps)

Model import example:

from rllte.hub.models import Atari

if __name__ == "__main__":
    model = Atari().load_models(
        agent="ppo",
        env_id="Pong-v5",
        seed=1,
        device="cuda"
    )
    print(model)

[yuanmingqi]

Environment: Envpool Procgen Games synchronous mode
Completed:

1. Proximal Policy Optimization (PPO)
   • ppo_procgen_bigfish (10 seeds, 25M steps)
   • ppo_procgen_bossfight (10 seeds, 25M steps)
   • ppo_procgen_caveflyer (10 seeds, 25M steps)
   • ppo_procgen_chaser (10 seeds, 25M steps)
   • ppo_procgen_climber (10 seeds, 25M steps)
   • ppo_procgen_coinrun (10 seeds, 25M steps)
   • ppo_procgen_dodgeball (10 seeds, 25M steps)
   • ppo_procgen_fruitbot (10 seeds, 25M steps)
   • ppo_procgen_heist (10 seeds, 25M steps)
   • ppo_procgen_jumper (10 seeds, 25M steps)
   • ppo_procgen_leaper (10 seeds, 25M steps)
   • ppo_procgen_maze (10 seeds, 25M steps)
   • ppo_procgen_miner (10 seeds, 25M steps)
   • ppo_procgen_ninja (10 seeds, 25M steps)
   • ppo_procgen_plunder (10 seeds, 25M steps)
   • ppo_procgen_starpilot (10 seeds, 25M steps)

Model import examle:

from rllte.hub.models import Procgen

if __name__ == "__main__":
    model = Procgen().load_models(
        agent="ppo",
        env_id="bigfish",
        seed=1,
        device="cuda"
    )
    print(model)