* feature(nyz): add hybrid ppo, unify action_space field and use dict type mu sigma

* polish(nyz): polish ppo config continous field, move to action_space field

* fix(nyz): fix ppo action_space field compatibility bug

* fix(nyz): fix ppg/sac/cql action_space field compatibility bug

* demo(nyz): update gym hybrid hppo config

* polish(pu): polish hppo hyper-para, use tanh and fixed sigma 0.3 in actor_action_args, use clamp [0,1] and [-1,1] for acceleration_value and rotation_value correspondingly after sample from the pi distri. in collect phase

* polish(pu):polish as review

* polish(pu): polish hppo config

* polish(pu): entropy weight=0.03 performs best empirically

* fix(nyz): fix unittest compatibility bugs

* polish(nyz): remove atari env unused print(ci skip)
Co-authored-by: Npuyuan1996 <2402552459@qq.com>

* add apple key to door treasure and polish

* add test, revise reward, build four envs

* add 7x7-1 ADTKT

* test rnd

* fix mz config

* fix config

* feature(pu): fix r2d2, add beta to actor

* feature(pu): add ngu-dev

* fix(pu): fix r2d2

* fix(puyuan): fix r2d2

* feature(puyuan): add minigrid r2d2 config

* polish minigrid config

* dev-ngu

* feature(pu): add action and reward as inputs of q network

* feature(pu): add episodic reward model

* feature(pu): add episodic reward model, modify r2d2 and collector for ngu

* fix(pu): recover files that were changed by mistake

* fix(pu): fix tblogger cnt bug

* add_dqfd

* Is_expert to is_expert

* fix(pu): fix r2d2 bug

* fix(pu): fix beta index to gamma bug

* fix(pu): fix numerical stability problem

* style(pu): flake8 format

* fix(pu): fix rnd reward model train times

* polish(pu): polish r2d2 reset problem

* fix(pu): fix episodic reward normalize bug

* polish(pu): polish config params and episodic_reward init value

* modify according to the last commnets

* value_gamma;done;marginloss;sqil适配

* feature(pu): add r2d3 algorithm and config of lunarlander and pong

* fix(pu): fix demo path bug

* fix(pu): fix cuda bug at function get_gae in adder.py

* feature(pu): add pong r2d2 config

* polish(pu): r2d2 uses the mixture priority, episodic_reward transforms to mean 0 std1

* polish(pu): polish r2d2 config

* test(pu): test cuda compatiality of dqfd_nstep_td_error in r2d3

* polish(pu): polish config

* polish(pu): polish config and annotation

* fix(pu): fix r2d2 target net update bug and done bug

* polish(pu): polish pong r2d2 config and add montezuma r2d2 config

* polish(pu): add some logs for debugging in r2d2

* polish(pu): recover config deleted by mistake

* fix(pu): fix r2d3 config of lunarlander and pong

* fix(pu): fix the r2d2 bug in r2d3

* fix(pu): fix r2d3 cpu device bug in fun dqfd_nstep_td_error of td.py

* fix(pu): fix n_sample bug in serial_entry_r2d3

* polish(pu): polish minigrid r2d2 config

* fix(pu): add info dict of fourrooms doorkey in minigrid_env

* polish(pu): polish r2d2 config

* fix(pu): fix expert policy collect traj bug, now we use the argmax_sample wrapper

* fix(pu): fix r2d2 done and target update bug, polish config

* fix(pu): fix null_padding transition obs to zeros

* fix(pu): episodic_reward transform to [0,1]

* fix(pu): fix the value_gamma bug

* fix(pu): fix device bug in ngu_reward_model.py

* fix(pu): fix null_padding problem in rnd and episodic reward model

* polish(pu): polish config

* fix(pu): use the deepcopy train_data to add bonus reward

* polish(pu): add the operation of enlarging seq_length times to the last reward of the whole episode

* fix(pu): fix the episode length 1 bug and weight intrinsic reward bug

* feature(pu): add montezuma ngu config

* fix(pu): fix lunarlander ngu unroll_len to 998 so that the sequence length is equal to the max step 1000

* test(pu): episodic reward transforms to [0,1]

* fix(pu): fix r2d3 one-step rnn init bug and add r2d2_collect_traj

* fix(pu): fix r2d2_collect_traj.py

* feature(pu): add pong_r2d3_r2d2expert_config

* polish(pu): yapf format

* polish(pu): fix td.py conflict

* polish(pu): flake8 format

* polish(pu): add lambda_one_step_td key in dqfd error

* test(pu): set key lambda_one_step_td and lambda_supervised_loss as 0

* style(pu): yapf format

* style(pu): format

* polish(nyz): fix ngu detailed compatibility error

* fix(nyz): fix dqfd one_step td lambda bug

* fix(pu): fix test_acer and test_rnd compatibility error
Co-authored-by: NSwain <niuyazhe314@outlook.com>
Co-authored-by: NWill_Nie <nieyunpengwill@hotmail.com>