IsaacGymEnvs-BallBalance-PPO
Trained agent for NVIDIA Isaac Gym Preview environments.
- Task: BallBalance
- Agent: PPO
Usage (with skrl)
Note: Visit the skrl Examples section to access the scripts.
PyTorch
from skrl.utils.huggingface import download_model_from_huggingface # assuming that there is an agent named `agent` path = download_model_from_huggingface("skrl/IsaacGymEnvs-BallBalance-PPO", filename="agent.pt") agent.load(path)
JAX
from skrl.utils.huggingface import download_model_from_huggingface # assuming that there is an agent named `agent` path = download_model_from_huggingface("skrl/IsaacGymEnvs-BallBalance-PPO", filename="agent.pickle") agent.load(path)
Hyperparameters
Note: Undefined parameters keep their values by default.
# https://skrl.readthedocs.io/en/latest/api/agents/ppo.html#configuration-and-hyperparameters
cfg = PPO_DEFAULT_CONFIG.copy()
cfg["rollouts"] = 16 # memory_size
cfg["learning_epochs"] = 8
cfg["mini_batches"] = 8 # 16 * 4096 / 8192
cfg["discount_factor"] = 0.99
cfg["lambda"] = 0.95
cfg["learning_rate"] = 3e-4
cfg["learning_rate_scheduler"] = KLAdaptiveRL
cfg["learning_rate_scheduler_kwargs"] = {"kl_threshold": 0.008}
cfg["random_timesteps"] = 0
cfg["learning_starts"] = 0
cfg["grad_norm_clip"] = 1.0
cfg["ratio_clip"] = 0.2
cfg["value_clip"] = 0.2
cfg["clip_predicted_values"] = True
cfg["entropy_loss_scale"] = 0.0
cfg["value_loss_scale"] = 2.0
cfg["kl_threshold"] = 0
cfg["rewards_shaper"] = lambda rewards, timestep, timesteps: rewards * 0.1
cfg["state_preprocessor"] = RunningStandardScaler
cfg["state_preprocessor_kwargs"] = {"size": env.observation_space, "device": device}
cfg["value_preprocessor"] = RunningStandardScaler
cfg["value_preprocessor_kwargs"] = {"size": 1, "device": device}
Evaluation results
- Total reward (mean) on IsaacGymEnvs-BallBalanceself-reported298.89 +/- 27.4