Tools for single-embodiment, multiple-task, Reinforcement Learning
Project description
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:
- tools for slicing, transforming, and composing specs
- 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
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distribution
Hashes for dm_robotics_agentflow-0.5.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | ee2569e405d9f19da0ab9d6201da55c2d86791a73e27fe891cd3d2537e88d946 |
|
MD5 | 213735e2ed31337ffac119974b7a857b |
|
BLAKE2b-256 | c9d2bc316a1fdddea90031f0a884c483bc1b9aec7582309e088cc0b3782c7176 |