Engineai RL Workspace
输入“/”快速插入内容
Engineai RL Workspace
Engineai RL Workspace
用户8132
用户52412025年12月4日修改
Overview
EngineAI RL Workspace is a versatile and universal RL framework for legged robots, including environment, training and eval for RL algorithms. It empowers seamless development and evaluation of RL algorithms for legged robots with modular design, extensibility, and comprehensive tools.
Features
Easy to develop
•
Modular design
Divided into individual modules, not only ensures readability, but also ensures that editing a single module will not affect other modules
◦
Env Obs, Domain Rands, Goals, Rewards
◦
Algo Runners, Algos, Networks, Storage
•
Simplification
◦
Train and play are controlled by the same runner logic and do not need to be written repeatedly
◦
Various modules can inherit according to the needs of exp, reducing repetitive writing
◦
Algo's change does not require changing Env, only changing the way output is received from Env
•
Extensibility
◦
Custom Env
◦
Custom Network
◦
Custom Algo
◦
Custom Testers & Tester Loggers
◦
Custom Training Logger
Comprehensive tools
◦
Record Runs
▪
Record videos of training or playing
▪
Custom Env variables and record play data to facilitate Run comparison under controlled variables
▪
Save the code file during training for easy comparison between runs
◦
Track and Reproduce Runs
▪
Automatically restore saved code files during resume and play, avoiding the tedious process of searching for corresponding versions
▪
Save git info
◦
Create Runs from JSON
▪
Save the .json file for recording parameters and have a built-in tool to convert it to a .py file, which is convenient for modification according to a certain run
◦
Format conversion
▪
Convert .pt files to .onnx and .mnn
◦
Multi-GPU training
▪
Support multi-GPU for faster training