RLLTE: Long-Term Evolution Project of Reinforcement Learning

</div>

Inspired by the long-term evolution (LTE) standard project in telecommunications, aiming to provide development components for and standards for advancing RL research and applications. Beyond delivering top-notch algorithm implementations, RLLTE also serves as a toolkit for developing algorithms.

Why RLLTE?

🧬 Long-term evolution for providing latest algorithms and tricks;
🏞️ Complete ecosystem for task design, model training, evaluation, and deployment (TensorRT, CANN, ...);
🧱 Module-oriented design for complete decoupling of RL algorithms;
🚀 Optimized workflow for full hardware acceleration;
⚙️ Support custom environments and modules;
🖥️ Support multiple computing devices like GPU and NPU;
💾 Large number of reusable benchmarks (RLLTE Hub);
🤖 Large language model-empowered copilot (RLLTE Copilot).

⚠️ Since the construction of RLLTE Hub requires massive computing power, we have to upload the training datasets and model weights gradually. Progress report can be found in Issue#30.

See the project structure below:

For more detailed descriptions of these modules, see API Documentation.

Quick Start

Installation

with pip recommended

Open a terminal and install rllte with pip:

shell
conda create -n rllte python=3.8 # create an virtual environment
pip install rllte-core # basic installation
pip install rllte-core[envs] # for pre-defined environments

with git

Open a terminal and clone the repository from GitHub with git:

sh
git clone https://github.com/RLE-Foundation/rllte.git
pip install -e . # basic installation
pip install -e .[envs] # for pre-defined environments

For more detailed installation instruction, see Getting Started.

Fast Training with Built-in Algorithms

RLLTE provides implementations for well-recognized RL algorithms and simple interface for building applications.

On NVIDIA GPU

Suppose we want to use DrQ-v2 to solve a task of DeepMind Control Suite, and it suffices to write a train.py like:

python
# import `env` and `agent` module
from rllte.env import make_dmc_env 
from rllte.agent import DrQv2

if __name__ == "__main__":
    device = "cuda:0"
    # create env, `eval_env` is optional
    env = make_dmc_env(env_id="cartpole_balance", device=device)
    eval_env = make_dmc_env(env_id="cartpole_balance", device=device)
    # create agent
    agent = DrQv2(env=env, eval_env=eval_env, device=device, tag="drqv2_dmc_pixel")
    # start training
    agent.train(num_train_steps=500000, log_interval=1000)

Run train.py and you will see the following output:

On HUAWEI NPU

Similarly, if we want to train an agent on HUAWEI NPU, it suffices to replace cuda with npu:

python
device = "cuda:0" -> device = "npu:0"

Three Steps to Create Your RL Agent

Developers only need three steps to implement an RL algorithm with RLLTE. The following example illustrates how to write an Advantage Actor-Critic (A2C) agent to solve Atari games.

Firstly, select a prototype:
<details> <summary>Click to expand code</summary> ``` py from rllte.common.prototype import OnPolicyAgent ``` </details>

Secondly, select necessary modules to build the agent:

<details> <summary>Click to expand code</summary>

py
from rllte.xploit.encoder import MnihCnnEncoder
from rllte.xploit.policy import OnPolicySharedActorCritic
from rllte.xploit.storage import VanillaRolloutStorage
from rllte.xplore.distribution import Categorical

Run the .describe function of the selected policy and you will see the following output:

py
OnPolicySharedActorCritic.describe()
# Output:
# ================================================================================
# Name       : OnPolicySharedActorCritic
# Structure  : self.encoder (shared by actor and critic), self.actor, self.critic
# Forward    : obs -> self.encoder -> self.actor -> actions
#            : obs -> self.encoder -> self.critic -> values
#            : actions -> log_probs
# Optimizers : self.optimizers['opt'] -> (self.encoder, self.actor, self.critic)
# ================================================================================

This illustrates the structure of the policy and indicate the optimizable parts.

</details>

Thirdly, merge these modules and write an .update function:

<details> <summary>Click to expand code</summary>

py
from torch import nn
import torch as th

class A2C(OnPolicyAgent):
    def __init__(self, env, tag, seed, device, num_steps) -> None:
        super().__init__(env=env, tag=tag, seed=seed, device=device, num_steps=num_steps)
        # create modules
        encoder = MnihCnnEncoder(observation_space=env.observation_space, feature_dim=512)
        policy = OnPolicySharedActorCritic(observation_space=env.observation_space,
                                          action_space=env.action_space,
                                          feature_dim=512,
                                          opt_class=th.optim.Adam,
                                          opt_kwargs=dict(lr=2.5e-4, eps=1e-5),
                                          init_fn="xavier_uniform"
                                          )
        storage = VanillaRolloutStorage(observation_space=env.observation_space,
                                        action_space=env.action_space,
                                        device=device,
                                        storage_size=self.num_steps,
                                        num_envs=self.num_envs,
                                        batch_size=256
                                        )
        dist = Categorical()
        # set all the modules
        self.set(encoder=encoder, policy=policy, storage=storage, distribution=dist)
    
    def update(self):
        for _ in range(4):
            for batch in self.storage.sample():
                # evaluate the sampled actions
                new_values, new_log_probs, entropy = self.policy.evaluate_actions(obs=batch.observations, actions=batch.actions)
                # policy loss part
                policy_loss = - (batch.adv_targ * new_log_probs).mean()
                # value loss part
                value_loss = 0.5 * (new_values.flatten() - batch.returns).pow(2).mean()
                # update
                self.policy.optimizers['opt'].zero_grad(set_to_none=True)
                (value_loss * 0.5 + policy_loss - entropy * 0.01).backward()
                nn.utils.clip_grad_norm_(self.policy.parameters(), 0.5)
                self.policy.optimizers['opt'].step()

</details>

Finally, train the agent by
<details> <summary>Click to expand code</summary> ``` py from rllte.env import make_atari_env if __name__ == "__main__": device = "cuda" env = make_atari_env("PongNoFrameskip-v4", num_envs=8, seed=0, device=device) agent = A2C(env=env, tag="a2c_atari", seed=0, device=device, num_steps=128) agent.train(num_train_steps=10000000) ``` </details>

As shown in this example, only a few dozen lines of code are needed to create RL agents with RLLTE.

Algorithm Decoupling and Module Replacement

RLLTE allows developers to replace settled modules of implemented algorithms to make performance comparison and algorithm improvement, and both
built-in and custom modules are supported. Suppose we want to compare the effect of different encoders, it suffices to invoke the .set function:

py
from rllte.xploit.encoder import EspeholtResidualEncoder
encoder = EspeholtResidualEncoder(...)
agent.set(encoder=encoder)

RLLTE is an extremely open framework that allows developers to try anything. For more detailed tutorials, see Tutorials.

Function List (Part)

RL Agents

Type	Algo.	Box	Dis.	M.B.	M.D.	M.P.	NPU	💰	🔭
On-Policy	A2C	✔️	✔️	✔️	✔️	✔️	✔️	✔️	❌
On-Policy	PPO	✔️	✔️	✔️	✔️	✔️	✔️	✔️	❌
On-Policy	DrAC	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️
On-Policy	DAAC	✔️	✔️	✔️	✔️	✔️	✔️	✔️	❌
On-Policy	DrDAAC	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️
On-Policy	PPG	✔️	✔️	✔️	❌	✔️	✔️	✔️	❌
Off-Policy	DQN	✔️	❌	❌	❌	✔️	✔️	✔️	❌
Off-Policy	DDPG	✔️	❌	❌	❌	✔️	✔️	✔️	❌
Off-Policy	SAC	✔️	❌	❌	❌	✔️	✔️	✔️	❌
Off-Policy	SAC-Discrete	❌	✔️	❌	❌	✔️	✔️	✔️	❌
Off-Policy	TD3	✔️	❌	❌	❌	✔️	✔️	✔️	❌
Off-Policy	DrQ-v2	✔️	❌	❌	❌	❌	✔️	✔️	✔️
Distributed	IMPALA	✔️	✔️	❌	❌	✔️	❌	❌	❌

Dis., M.B., M.D.: Discrete, MultiBinary, and MultiDiscrete action space;

M.P.: Multi processing;

🐌: Developing;

💰: Support intrinsic reward shaping;

🔭: Support observation augmentation.

Intrinsic Reward Modules

Type	Modules
Count-based	PseudoCounts, RND, E3B
Curiosity-driven	ICM, GIRM, RIDE, Disagreement
Memory-based	NGU
Information theory-based	RE3, RISE, REVD

See Tutorials: Use Intrinsic Reward and Observation Augmentation for usage examples.

RLLTE Ecosystem

Explore the ecosystem of RLLTE to facilitate your project:

Hub: Fast training APIs and reusable benchmarks.
Evaluation: Reasonable and reliable metrics for algorithm evaluation.
Env: Packaged environments for fast invocation.
Deployment: Convenient APIs for model deployment.
Pre-training: Methods of pre-training in RL.
Copilot: Large language model-empowered copilot.

How To Contribute

Welcome to contribute to this project! Before you begin writing code, please read CONTRIBUTING.md for guide first.

Cite the Project

To cite this project in publications:

bibtex
@inproceedings{
yuan2025rllte,
title={{RLLTE}: Long-Term Evolution Project of Reinforcement Learning},
author={Mingqi Yuan and Zequn Zhang and Yang Xu and Jake Shihao Luo and Bo Li and Xin Jin and Wenjun Zeng},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2025}
}

Acknowledgment

This project is supported by The Hong Kong Polytechnic University, Eastern Institute for Advanced Study, and FLW-Foundation. EIAS HPC provides a GPU computing platform, and HUAWEI Ascend Community provides an NPU computing platform for our testing. Some code of this project is borrowed or inspired by several excellent projects, and we highly appreciate them. See ACKNOWLEDGMENT.md.