An open-source Gym-compatible environment specifically tailored for developing RL algorithms for autonomous driving. DriverGym provides access to more than 1000 hours of expert logged data and also supports reactive and data-driven agent behavior. The performance of an RL policy can be easily validated using an extensive and flexible closed-loop evaluation protocol. We also provide behavior cloning baselines using supervised learning and RL, trained in DriverGym.
A platform for running reproducible reinforcement learning experiments for customizable robotic reaching tasks. This self-contained and straightforward toolbox allows its users to quickly investigate and identify optimal training configurations.
The Github repository contains code, CAD files and a bill of materials for you to build the robot. You can watch `a video overview of the project here <https://www.youtube.com/watch?v=Y6FVBbqjR40>`_.
A solution to the second project of the Udacity deep reinforcement learning course.
It is an example of:
- wrapping single and multi-agent Unity environments to make them usable in Stable-Baselines3
- creating experimentation scripts which train and run A2C, PPO, TD3 and SAC models (a better choice for this one is https://github.com/DLR-RM/rl-baselines3-zoo)
- generating several pre-trained models which solve the reacher environment
SuperSuit contains easy to use wrappers for Gym (and multi-agent PettingZoo) environments to do all forms of common preprocessing (frame stacking, converting graphical observations to greyscale, max-and-skip for Atari, etc.). It also notably includes:
-Wrappers that apply lambda functions to observations, actions, or rewards with a single line of code.
-All wrappers can be used natively on vector environments, wrappers exist to Gym environments to vectorized environments and concatenate multiple vector environments together
-A wrapper is included that allows for using regular single agent RL libraries (e.g. stable baselines) to learn simple multi-agent PettingZoo environments, explained in this tutorial:
| Tutorial on multi-agent support in stable baselines: https://towardsdatascience.com/multi-agent-deep-reinforcement-learning-in-15-lines-of-code-using-pettingzoo-e0b963c0820b
Suite of RL environments focused on using a simulated tactile sensor as the primary source of observations. Sim-to-Real results across 4 out of 5 proposed envs.
RLeXplore is a set of implementations of intrinsic reward driven-exploration approaches in reinforcement learning using PyTorch, which can be deployed in arbitrary algorithms in a plug-and-play manner. In particular, RLeXplore is designed to be well compatible with Stable-Baselines3, providing more stable exploration benchmarks.
A simple library for pink noise exploration with deterministic (DDPG / TD3) and stochastic (SAC) off-policy algorithms. Pink noise has been shown to work better than uncorrelated Gaussian noise (the default choice) and Ornstein-Uhlenbeck noise on a range of continuous control benchmark tasks. This library is designed to work with Stable Baselines3.
| Authors: Onno Eberhard, Jakob Hollenstein, Cristina Pinneri, Georg Martius
A Deep Reinforcement Learning Open-Source Toolkit for Network Slice Placement (NSP).
NSP is the problem of deciding which physical servers in a network should host the virtual network functions (VNFs) that make up a network slice, as well as managing the mapping of the virtual links between the VNFs onto the physical infrastructure.
It is a complex optimization problem, as it involves considering the requirements of the network slice and the available resources on the physical network.
The goal is generally to maximize the utilization of the physical resources while ensuring that the network slices meet their performance requirements.
The toolkit includes a customizable simulation environments, as well as some ready-to-use demos for training
intelligent agents to perform network slice placement.
| Paper: **under review** (citation instructions on the project's README.md) -> see this Master's Thesis for the moment: https://etd.adm.unipi.it/theses/available/etd-01182023-110038/unrestricted/Tesi_magistrale_Pasquali_Alex.pdf
Meta-RL framework to optimize reservoir-like neural structures (special kind of RNNs), and integrate them to RL agents to improve their training.
It enables solving environments involving partial observability or locomotion (e.g MuJoCo), and optimizing reservoirs that can generalize to unseen tasks.
These environments are dedicated to train efficient agents that can plan and forecast bipedal robot footsteps in order to go to a target location possibly avoiding obstacles. They are designed to be used with Reinforcement Learning (RL) algorithms.
Real world experiments were conducted during RoboCup competitions on the Sigmaban robot, a small-sized humanoid designed by the *Rhoban Team*.
