Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/lorenmt/minimal-isaac-gym
A Minimal Example of Isaac Gym with DQN and PPO.
https://github.com/lorenmt/minimal-isaac-gym
dqn isaac-gym ppo pytorch
Last synced: 7 days ago
JSON representation
A Minimal Example of Isaac Gym with DQN and PPO.
- Host: GitHub
- URL: https://github.com/lorenmt/minimal-isaac-gym
- Owner: lorenmt
- Created: 2022-03-14T11:54:23.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2023-05-11T13:26:16.000Z (over 1 year ago)
- Last Synced: 2024-08-02T11:22:31.496Z (3 months ago)
- Topics: dqn, isaac-gym, ppo, pytorch
- Language: Python
- Homepage:
- Size: 12.7 KB
- Stars: 87
- Watchers: 2
- Forks: 13
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-isaac-gym - minimal-isaac-gym
README
# Minimal Isaac Gym
This repository provides a minimal example of NVIDIA's [Isaac Gym](https://developer.nvidia.com/isaac-gym), to assist other researchers like me to quickly understand the code structure, to be able to design fully customised large-scale reinforcement learning experiments.The example is based on the official implementation from the Isaac Gym's [Benchmark Experiments](https://github.com/NVIDIA-Omniverse/IsaacGymEnvs), for which we have followed a similar implementation on [Cartpole](https://github.com/NVIDIA-Omniverse/IsaacGymEnvs/blob/main/isaacgymenvs/tasks/cartpole.py), but with a minimal number of lines of code aiming for maximal readability, and without using any third-party RL frameworks.
**Note**: The current implementation is based on Isaac Gym Preview Version 3, with the support for two RL algorithms: *DQN* and *PPO* (both continuous and discrete version). PPO seems to be the default RL algorithm for Isaac Gym from the recent works of [Learning to walk](https://arxiv.org/abs/2109.11978) and [Object Re-orientation](https://arxiv.org/abs/2111.03043), since it only requires on-policy training data and therefore to make it a much simpler implementation coupled with Isaac Gym's APIs.
*Both DQN and PPO are expected to converge under 1 minute.*
## Usage
Simply run `python trainer.py --method {dqn; ppo, ppo_d}`.## Disclaimer
I am also very new to Isaac Gym, and I cannot guarantee my implementation is absolutely correct. If you have found anything unusual or unclear that can be improved, PR or Issues are highly welcomed.