gymnasium.Env inheritance

reset() return tuple

step() return 5-tuple

Image dtype uint8

Channel-first image shape

check_env called

CnnPolicy used

Discrete space starts at 0

self.np_random used

Image pixel bounds [0, 255]

make_vec_env used

gradient_steps=-1

eval_freq adjusted for n_envs

save_freq adjusted for n_envs

Separate eval environment

Multiple callbacks chained

No DDPG used

model.save() called

SAC or TD3 for continuous control

VecEnv type used

Static model load

evaluate_policy used

deterministic=True in evaluate_policy

VecEnv step() returns 4-tuple

VecEnv reset() returns obs only

terminal_observation accessed correctly

Replay buffer caveat noted

Installation uses uv pip install

results.json written

No DDPG used

HerReplayBuffer class used

SAC or TD3 base algorithm

MultiInputPolicy used

Goal obs structure keys

compute_reward method defined

goal_selection_strategy set

n_sampled_goal set

gymnasium.Env inheritance

check_env called

model.save() called

No DDPG used

uv pip install

VecNormalize applied

VecNormalize stats saved separately

VecNormalize training disabled on load

norm_reward disabled on load

linear_schedule returns progress_remaining * value

Schedule passed to model

VecNormalize.load used

model.save() called

Separate eval environment

evaluate_policy called

uv pip install used

BaseCallback subclass

_on_step returns bool

self.num_timesteps used

self.model or self.training_env used

self.logger.record used

StopTrainingOnRewardThreshold via callback_on_new_best

EvalCallback used

Multiple callbacks combined

Callback log file or printed output

uv pip install used

No DDPG used