wrapper problem under simple_speaker_listener environment

[CaptainYin]

when using simple_speaker_listener environment, the observation dimension of speaker and lisener is not the same, so there is a problem when wrapper the env by env_wrappers.py

execute this shell

`#!/bin/sh
env="MPE"
scenario="simple_speaker_listener"
#"simple_speaker_listener" "simple_spread"
num_landmarks=3
num_agents=2
algo="rmaddpg"
exp="debug"
seed_max=1

echo "env is ${env}, scenario is ${scenario}, algo is ${algo}, exp is ${exp}, max seed is ${seed_max}"

for seed in $(seq ${seed_max}); do
echo "seed is ${seed}:"
CUDA_VISIBLE_DEVICES=0 python3 train/train_mpe.py --env_name ${env} --n_rollout_threads 1 --algorithm_name ${algo} --experiment_name ${exp} --scenario_name ${scenario} --num_agents ${num_agents} --num_landmarks ${num_landmarks} --seed ${seed} --episode_length 25 --actor_train_interval_step 1 --tau 0.005 --lr 7e-4 --num_env_steps 10000000 --use_reward_normalization --share_policy
echo "training is done!"
done
outputenv is MPE, scenario is simple_speaker_listener, algo is rmaddpg, exp is debug, max seed is 1
seed is 1:
choose to use gpu...
Namespace(algorithm_name='rmaddpg', experiment_name='debug', seed=1, cuda=True, cuda_deterministic=True, n_training_threads=1, n_rollout_threads=1, n_eval_rollout_threads=1, num_env_steps=10000000, use_wandb=False, user_name='chris_2024', env_name='MPE', use_obs_instead_of_state=False, episode_length=25, buffer_size=5000, use_reward_normalization=True, use_popart=False, popart_update_interval_step=2, use_per=False, per_nu=0.9, per_alpha=0.6, per_eps=1e-06, per_beta_start=0.4, use_centralized_Q=True, share_policy=False, hidden_size=64, layer_N=1, use_ReLU=True, use_feature_normalization=True, use_orthogonal=True, gain=0.01, use_conv1d=False, stacked_frames=1, prev_act_inp=False, use_rnn_layer=True, use_naive_recurrent_policy=True, recurrent_N=1, data_chunk_length=80, burn_in_time=0, attn=False, attn_N=1, attn_size=64, attn_heads=4, dropout=0.0, use_average_pool=True, use_cat_self=True, lr=0.0007, opti_eps=1e-05, weight_decay=0, batch_size=32, gamma=0.99, use_max_grad_norm=True, max_grad_norm=10.0, use_huber_loss=False, huber_delta=10.0, use_soft_update=True, tau=0.005, hard_update_interval_episode=200, hard_update_interval=200, target_action_noise_std=0.2, alpha=1.0, target_entropy_coef=0.5, automatic_entropy_tune=True, use_double_q=True, hypernet_layers=2, mixer_hidden_dim=32, hypernet_hidden_dim=64, num_random_episodes=5, epsilon_start=1.0, epsilon_finish=0.05, epsilon_anneal_time=50000, act_noise_std=0.1, actor_train_interval_step=1, train_interval_episode=1, train_interval=100, use_value_active_masks=False, use_eval=True, eval_interval=10000, num_eval_episodes=32, save_interval=100000, log_interval=1000, model_dir=None, scenario_name='simple_speaker_listener', num_landmarks=3, num_agents=2, use_same_share_obs=True)
warm up...
Traceback (most recent call last):
File "/project/off-policy-release/offpolicy/scripts/train/train_mpe.py", line 193, in
main(sys.argv[1:])
File "/project/off-policy-release/offpolicy/scripts/train/train_mpe.py", line 176, in main
runner = Runner(config=config)
File "/project/off-policy-release/offpolicy/runner/rnn/mpe_runner.py", line 15, in init
self.warmup(num_warmup_episodes)
File "/project/off-policy-release/offpolicy/runner/rnn/base_runner.py", line 221, in warmup
env_info = self.collecter(explore=True, training_episode=False, warmup=True)
File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/project/off-policy-release/offpolicy/runner/rnn/mpe_runner.py", line 145, in separated_collect_rollout
obs = env.reset()
File "/project/off-policy-release/offpolicy/envs/env_wrappers.py", line 439, in reset
return np.array(obs)
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (1, 2) + inhomogeneous part.
training is done!`