Skip to content

AllenAct Tutorials#

Note The provided commands to execute these tutorials assume you have installed the full library and the specific requirements for each used plugin.

We provide several tutorials to help ramp up researchers to the field of Embodied-AI as well as to the AllenAct framework.

MiniGridEmptyRandom5x5 task example

We train an agent to complete the MiniGrid-Empty-Random-5x5-v0 task within the MiniGrid environment.

This tutorial presents:

  • Writing an experiment configuration file with a simple training pipeline from scratch.
  • Using one of the supported environments with minimal user effort.
  • Training, validation and testing your experiment from the command line.

Follow the tutorial here.

PointNav in RoboTHOR#

RoboTHOR Robot

We train an agent on the Point Navigation task within the RoboTHOR Embodied-AI environment.

This tutorial presents:

  • The basics of the Point Navigation task, a common task in Embodied AI
  • Using an external dataset
  • Writing an experiment configuration file with a simple training pipeline from scratch.
  • Use one of the supported environments with minimal user effort.
  • Train, validate and test your experiment from the command line.
  • Testing a pre-trained model

Follow the tutorial here.

Swapping in a new environment#

Environment Transfer

This tutorial demonstrates how easy it is modify the experiment config created in the RoboTHOR PointNav tutorial to work with the iTHOR and Habitat environments.

Follow the tutorial here.

Using a pretrained model#

Pretrained inference

This tutorial shows how to run inference on one or more checkpoints of a pretrained model and generate visualizations of different types.

Follow the tutorial here.

Off-policy training#

This tutorial shows how to train an Actor using an off-policy dataset with expert actions.

Follow the tutorial here.

OpenAI gym for continuous control#

gym task example

We train an agent to complete the LunarLanderContinuous-v2 task from OpenAI gym.

This tutorial presents:

  • A gym plugin fopr AllenAct.
  • A continuous control example with multiple actors using PPO.

Follow the tutorial here.

Multi-node training for RoboTHOR ObjectNav#

training speedup

We train an agent to navigate to an object in a fraction of the time required for training in one node by distributing training across multiple nodes.

This tutorial presents:

  1. The AllenAct API for training across multiple nodes, as well as experimental scripts for distributed configuration, training start and termination, and remote command execution.
  2. The introduction of the headless mode for AI2-THOR in AllenaAct.

Follow the tutorial here.