AgentFlow: A Modular Toolkit for Scalable RL Research

Overview

AgentFlow is a library for composing Reinforcement-Learning agents. The core features that AgentFlow provides are:

1. tools for slicing, transforming, and composing specs
2. tools for encapsulating and composing RL-tasks.

Unlike the standard RL setup, which assumes a single environment and an agent, AgentFlow is designed for the single-embodiment, multiple-task regime. This was motivated by the robotics use-case, which frequently requires training RL modules for various skills, and then composing them (possibly with non-learned controllers too).

Instead of having to implement a separate RL environment for each skill and combine them ad hoc, with AgentFlow you can define one or more SubTasks which modify a timestep from a single top-level environment, e.g. adding observations and defining rewards, or isolating a particular sub-system of the environment, such as a robot arm.

You then compose SubTasks with regular RL-agents to form modules, and use a set of graph-building operators to define the flow of these modules over time (hence the name AgentFlow).

The graph-building step is entirely optional, and is intended only for use-cases that require something like a (possibly learnable, possibly stochastic) state-machine.

Components

Control Flow

Examples