Environment Wrappers

NormalizeWrapperEnv

Normalizes observations and rewards using running statistics.

env = NormalizeWrapperEnv(env;
    normalize_obs = true,
    normalize_reward = true,
    clip_obs = 10.0f0,
    clip_reward = 10.0f0,
)

# Save/load normalization stats
save_normalization_stats(env, "norm_stats.jld2")
load_normalization_stats!(env, "norm_stats.jld2")

# Toggle training mode (stops updating stats)
set_training(env, false)

ScalingWrapperEnv

Scales observations and actions to normalized ranges.

env = ScalingWrapperEnv(env)

Warning

ScalingWrapperEnv requires finite bounds on observation and action spaces. It will not work with spaces containing Inf or -Inf bounds.

MonitorWrapperEnv

Tracks episode statistics.

env = MonitorWrapperEnv(env)
# Access stats via get_info(env)

Wrapper Utilities

# Check if wrapped
is_wrapper(env)

# Unwrap one layer
inner_env = unwrap(env)

# Unwrap all layers
base_env = unwrap_all(env)