Roadmap for DI-engine

[PaParaZz1]

This issue will keep tracking of DI-engine's updates of next few versions:

Future

• feature(wrh): add EDT code #808
• feature(xrk): add q-transformer #783
• feature(zc): add MetaDiffuser and prompt-dt #771
• feature(zjow): add envpool new pipeline #753
• feature(whl): add rlhf pipeline. #748
• refactor(gry): refactor reward model #636
• feature(whl): add PC+MCTS code #603
• feature(wgt): enable DI using torch-rpc to support GPU-p2p and RDMA-rpc #562
• feature(zms): add new league middlewares and other models and tools.  #458

v0.5.4

(Expected to release in Dec 2025)

• BUG 代码有问题 #875
• [Secuirty] Remote Code Execution by Pickle Deserialization via ding/framework/parallel/_handle_message() in opendilab/DI-engine #874
• dizoo中的metadrive示例metadrive_onppo_config.py跑不起来 #871
• di-engine 0.5.3 requires numpy<2,>=1.18.0 #869
• How to increase the hybrid action space in multi-agent environments #868
• fix(pu): fix noise layer's usage based on the original paper #866
• feature(pu): adapt to unizero-multitask ddp, and adapt ppo to support jericho config #860
• feature(nyz&dcy): add LLM/VLM reward model #859
• feature(nyz&dcy): add LLM/VLM RLHF loss (PPO/GRPO/RLOO) #857
• feature(wqj): add vllm collector #856
• polish(pu): delete unused enable_fast_timestep argument #855
• feature(nyz): add rlhf dataset #854
• AttributeError: 'MultiDiscrete' object has no attribute 'low' #852
• Noisy Net Issue #850
• LSTM Layer normalization #849
• feature(zjow): add Implicit Q-Learning #821

v0.5.3

• Bug in reset method #846
• Reset env_info #845
• Priority Experience Replay Bug #844
• feature(xyy):add HPT model to implement PolicyStem+DuelingHead #841
• conda安装di-engine，无法与python 3.9兼容 #838
• feature(pu): add resume_training option to allow the envstep and train_iter resume seamlessly #835
• feature(pu): add pistonball_env, its unittest and qmix config #833
• polish(TairanMK): update trading env #831
• polish(mark): add hybrid action space support to ActionNoiseWrapper #829
• feature(whl): add AWR algorithm. #828
• No Hidden Size List for ContinuousQAC? #826
• Observation shape in the custom marl environment #823
• feature(nyz): adapt DingEnvWrapper to gymnasium #817
• KeyError: 'obs_shape' #812
• feature(zym): update ppo config to support discrete action space #809

v0.5.2

• style(nyz): relax flask requirement #811
• feature(wrh): add taxi env latest version and dqn config #807
• style(hus): add new badge (hellogithub) in readme #805
• polish(zym): optimize ppo continuous act #801
• feature(wrh): add taxi env #799
• cannot run GTrXL demo since v0.5.0 #796
• doc(hus): update discord link and badge in readme #795
• docker内运行lunarlander_dqn_deploy失败 #793
• feature(nyz): add GPU utils #788
• fix(zjow): fix complex obs demo for ppo pipeline #786
• env(rjy): add ising model env #782
• feature(xrk): add new env named Flozen Lake and DQN algorithm.  #781
• feature(ooo): add deprecated function decorator #778
• fix(eltociear): typo in config.py #776
• feature(nyz): add MADDPG pettingzoo example #774

v0.5.1

• feature(nyz): add MADDPG pettingzoo example #774
• polish(pu): polish comments in a2c/bcq/fqf/ibc policy #768
• polish(pu): polish NGU atari configs #767
• polish(rjy): polish pg/iqn/edac policy doc #764
• doc(zjow): polish the notation of classes and functions in torch_utils and utils #763
• doc(rjy): polish d4pg/ppg/qrdqn policy doc #762
• fix(pu): fix hppo entropy_weight to avoid nan error in log_prob #761
• doc(zjow): add API doc for ding agent #758
• feature(zjow): add qgpo policy for new DI-engine pipeline #757
• polish(rjy): polish the comments of collate_fn/profiler_helper/metric #755
• feature(luyd): fix dt new pipeline of mujoco #754
• polish(rjy): polish comments in normalizer_helper and lock_helper #752
• doc(whl): polish doc for loss, compression helper and bfs helper. #747
• polish(pu): polish comments and styles in files within torch_utils/network #745
• feature(cy): add dreamerV3 + MiniGrid code #725
• feature(rjy): add HAPPO algorithm #717

v0.5.0

• polish(zc): change PD config name #749
• polish(pu): polish comments in env_wrappers.py and ding_env_wrapper.py #742
• doc(zjow): polish ding model common/template note #741
• polish(rjy): polish comments in wqmix/ngu/pg model #739
• doc(whl): polish doc for data_helper, model_helper, parameter, metric, math_helper, backend_helper #738
• env(zjow): add Huggingface model card support for ppof envs #737
• polish(rjy): polish comments of qmix/pdqn/mavac #736
• feature(luyd): add collector logging in new pipeline #735
• doc(whl): add code doc for LT,DT,PC,BC models #734
• doc(zjow): update README.md and Colab demo #733
• polish(nyz): polish dqn and ppo comments #732
• feature(cy): polish anytrading #731
• feature(zjow): polish ppof agent code for opendilab huggingface #730
• fix(zjow): fix typo for QAC class #729
• test(luyd): add model test code #728
• style(eltociear): fix typo in optimizer_helper.py #726
• doc(zjow): polish rl_utils doc #724
• style(nyz): polish model template comments #722
• how to use logger #715
• fix(luyd): fix new pipeline impala in Lunarlander and Atari env #713
• feature(lxy): add dropout layers to dqn #712
• Error occurs when running in parallel #709
• feature(zc): add plan diffuser #700
• feature(whl): add tabmwp env and prompt pg policy #667
• feature(zjow): add new pipeline agent sac/ddpg/a2c #637

v0.4.9

• fix(lixuelin): to_ndarray fails to assign dtype for scalars #708
• feature(whl): add example of dqn eval #706
• fix(zyz): fix type spell error #704
• feature(pu): add three variants of Bilinear classes and a FiLM class #703
• feature(zjow): add middleware for ape-x structure pipeline #696
• doc(zjow): update README.md with openxlab badge #695
• refactor(lyd): refactor dt_policy in new pipeline and add img input support #693
• enable_save_figure() has no detailed document, and its comment is wrong. #688
• polish(nyz): polish offpolicy RL multi-gpu DDP training #679
• feature(nyz): fix py37 macos ci bug and update default pytorch to 1.12.1 #678
• feature(cxy): add cliffwalking env #677
• feature(cy): add tensor stream merging tools #673
• polish(nyz): simplify requirements #672
• feature(zp): add dreamerv3 algorithm #652
• feature(zc): add bcq algorithm #640

v0.4.8

• HELP: failed to run gym_hybrid_pdqn_config..py #664
• feature(nyz): add MAPPO/MASAC task example #661
• polish(pu): add LN and GN norm_type support in ResBlock #660
• fix(zjow): update td3bc d4rl config #659
• fix(pu): fix last_linear_layer_weight_bias_init_zero in MLP and add its unittest #650
• 运行DI-Zoo里面写好的config报错 #646
• feature(zjow): add PPO demo for complex env observation #644
• feature(zc): add EDAC and modify config of td3bc #639
• Missing of Regularization of GCL #627
• feature(wgt): add barrier middleware #570

v0.4.7

• fix(nyz): fix confusing shallow copy operation bug about next_obs #641
• feature(wgt): add torch-rpc fix dockerfile #628
• feature(whl): add gpt utils #625
• polish(gry): polish reward model and td error #624
• refactor(nyz): remove policy cfg sub fields requirements #620
• polish(gry): polish dqn config #611
• Tag v0.4.6 use __version__ 0.4.5 #609
• feature(nyz): add PPOF ch4 reward demo suuport #608
• feature(lxy): add popart & value rescale & symlog to ppof #605
• style(elt): fix typo in time_helper.py #602
• Document of SAC default config need to be updated #601
• fix(psharold): unsqueeze action_args in PDQN when shape is 1 #599
• fix(nyz): update ptz to latest version #597
• The "serial_pipeline" method has a small bug #592
• feature(lxy): modify ppof rewardclip and add atari config #589
• feature(gry): add MDQN algorithm #590
• Problem with DQfD doc #585
• MPDQN throw a RuntimeError #583
• feature(zjow): add wandb logger features; fix relative bugs for wandb online logger #579
• Examples for using Imitation Learning / Offline RL Algorithms #576
• fix(lisong): fix icm/rnd+onppo config bugs and app_key env bugs #564
• feature(whl): add PC algorithm #514
• feature(rjy): add dmc2gym+sac baseline both in state input and pixel input #451

v0.4.6

• feature(nyz): add ppof ch3 demo #581
• feature(nyz): setup evogym docker #580
• polish(whl): Update cartpole config #578
• feature(gry): add acrobot env and dqn config #577
• feature(zc): add carracing in box2d #575
• feature(zt): add metadrive-simulator env and related onppo config #574
• feature(lxy): add procedure cloning model #573
• style(eltociear): fix typo in contrastive_loss.py #572
• polish(nyz): polish example demos #568
• feature(nyz): add PPOF new interface support #567
• fix(nyz): fix py38 unittest bugs #565
• feature(nyz): add new gym hybrid viz #563
• feature(cy): add BDQ algorithm #558

v0.4.5

• feature(nyz): add policy gradient algo implementation #544
• feature(zjow): add load and save method for replaybuffer #542
• feature(wyh): madqn algorithm #540
• feature(whl): add more DingEnvWrapper example #525
• feature(nyz): add evaluator more info viz support #538
• env(zjow): add env gym_pybullet_drones #526
• env(lisong): add beergame supply chain optimization env #512
• ModuleNotFoundError: No module named 'ding' #535
• feature(zjow): add trackback log for subprocess env manager #534
• feature(nyz): add MADDPG algo #550
• Continuous action example for IMPALA #543
• polish(zjow): rename 'eval reward' -> 'episode return' #536
• feature(nyz): add new middleware distributed demo #321