Gym Fetch Stack

Mujoco Block Stacking Gym Reinforcement Learning Environments.
(Modified from OpenAI Robotics Gym Environments)

These environments are made for use with DDPG with Curiosity Driven Exploration and Multi-Criteria Hindsight Experience Replay

Setup

(You need a Mujoco License. Follow the instructions to set up Mujoco here.)

In the gym_fetch_stack root dir, use

pip install -e .

In your python code, use:

import gym
import gym_fetch_stack
env = gym.make("FetchStack2Stage3-v1")

Curriculum Stages

We use three curriculum stages to train an agent to stack blocks:

• Stage 1: basic manipulation tasks without having to create stacks
• Stage 2: stacking blocks where the environment is initialized at various stages of completion
• Stage 3: stacking blocks where all blocks all initialized away from their target locations

A video example of the different stages can found here.

Environments Available:

Incremental Rewards (Sparse reward for each correctly placed block):

(FetchStacki has i blocks in the environment to stack)

• FetchStack2Stage1-v1

• FetchStack3Stage1-v1

• FetchStack4Stage1-v1

• FetchStack2Stage2-v1

• FetchStack3Stage2-v1

• FetchStack4Stage2-v1

• FetchStack2Stage3-v1

• FetchStack3Stage3-v1

• FetchStack4Stage3-v1

Binary Rewards (Single Sparse reward for completed stack):

• FetchStack2SparseStage1-v1

• FetchStack3SparseStage1-v1

• FetchStack4SparseStage1-v1

• FetchStack2SparseStage2-v1

• FetchStack3SparseStage2-v1

• FetchStack4SparseStage2-v1

• FetchStack2SparseStage3-v1

• FetchStack3SparseStage3-v1

• FetchStack4SparseStage3-v1