RoboFeederEnv#

This document describes the RoboFeeder environments available in the Gym4ReaL library for robotic grasping tasks using reinforcement learning. These environments are designed for use with the MuJoCo simulator and the Gymnasium interface, enabling efficient training and evaluation of robotic agents.

Overview#

The RoboFeeder environments simulate a robotic arm tasked with picking objects from a box. The simulation includes:

  • A Staubli TX2-60 robot with a simple gripper.

  • A box containing multiple objects to pick.

  • A virtual overhead camera for observation.

  • Optional object detection features.

The environments are compatible with GPU acceleration for training.

Conda Usage#

conda create -n env-name python=3.12

Installation#

To install the general and environment-specific requirements,if gpus are not available run:

pip install -r requirements.txt
pip install -r gym4real/envs/robofeeder/requirements.txt

Otherwise:

pip install -r requirements.txt
pip install -r gym4real/envs/robofeeder/requirements-gpus.txt

RoboFeeder Environments#

The following environments are implemented in Gym4ReaL:

RoboFeeder-v0#

  • Observation Space: Cropped RGB image from the overhead camera, with the help of a pretrianed object dection neural network.

  • Action Space: Cartesian coordinates for the robot end-effector.

  • Goal: Pick an object based on visual input.

RoboFeeder-v1#

  • Observation Space: Full RGB image from the overhead camera.

  • Action Space: Cartesian coordinates for the robot end-effector.

  • Goal: Pick an object based on visual input.

RoboFeeder-v2#

  • Observation Space: List of segmented images for all objects in the box.

  • Action Space: Index of the selected object and pick coordinates.

  • Goal: Select and pick graspable object from the robot workspace.

Usage#

To use these environments:

  1. Download the Gym4ReaL library and install its dependencies.

  2. Register and instantiate the desired RoboFeeder environment using Gym’s API.

  3. Train and evaluate your custom or off-the-shelf RL agent implementation.

Example:

import gymnasium as gym

env = gym.make('gym4real/robofeeder-picking-v0')
obs,info = env.reset()
done = False
while not done:
  action = env.action_space.sample()
  obs, reward, done, info = env.step(action)

Configuration#

Environment parameters (e.g., camera resolution, robot workspace limits, number of objects) can be set via the configuration.yaml file in the project root.

Reproducibility#

In order to reproduce the results, open the notebooks in examples/robofeeder/benchmarks folder and run the whole notebook depending on the selected environment.

To reproduce the results from scratch, launch this command from the main directory selecting the environment and the training parameters:

git clone https://github.com/Daveonwave/gym4ReaL.git
cd gym4Real
python gym4real/algorithms/robofeeder/ppo.py --env gym4real/robofeeder-planning --episodes 1000 --batch-size 64 --learning-rate 0.0003

You can adjust the arguments (--env, --episodes, --batch-size, --learning-rate, etc.) as needed. See ppo.py --help for all available options.