Below is the list of implemented environments. Each environment (observation variant) has two alternatives, Task-Obs-vX and Task-Obs-Gazebo-vX (omitted from the table). Here, Task-Obs-vX implements the logic of the environment and can be used on real robots, whereas Task-Obs-Gazebo-vX combines this logic with the simulation environment inside Gazebo. Robots should be interchangeable for most parts, with some limitations (e.g. GraspPlanetary task requires a mobile manipulator to randomize the environment fully).

If you are interested in configuring these environments, first take a look at the list of their parameters inside Gym registration and then at their individual source code.

Reach the end-effector goal.	Grasp and lift a random object.	Grasp and lift a Moon rock.
Reach-v0 (state obs)	Grasp-v0 (state obs)	GraspPlanetary-v0 (state obs)
—	—	GraspPlanetary-MonoImage-v0
Reach-ColorImage-v0	—	GraspPlanetary-ColorImage-v0
Reach-DepthImage-v0	—	GraspPlanetary-DepthImage-v0
—	—	GraspPlanetary-DepthImageWithIntensity-v0
—	—	GraspPlanetary-DepthImageWithColor-v0
Reach-Octree-v0	Grasp-Octree-v0	GraspPlanetary-Octree-v0
Reach-OctreeWithIntensity-v0	Grasp-OctreeWithIntensity-v0	GraspPlanetary-OctreeWithIntensity-v0
Reach-OctreeWithColor-v0	Grasp-OctreeWithColor-v0	GraspPlanetary-OctreeWithColor-v0

By default, Grasp and GraspPlanetary tasks utilize GraspCurriculum that shapes their reward function and environment difficulty.

To facilitate the sim-to-real transfer of trained agents, simulation environments introduce domain randomization with the aim of improving the generalization of learned policies. This randomization is accomplished via ManipulationGazeboEnvRandomizer that populates the virtual world and enables randomizing of several properties at each reset of the environment. As this randomizer is configurable with numerous parameters, please take a look at the source code to see what environments you can create.

Examples of domain randomization for the Grasp task.

Examples of domain randomization for the GraspPlanetary task.

Model Datasets

Simulation environments in this repository can utilize datasets of any SDF models, e.g. models from Fuel. By default, the Grasp task uses Google Scanned Objects collection together with a set of PBR textures pointed to by TEXTURE_DIRS environment variable. On the contrary, the GraspPlanetary task employs custom models that are procedurally generated via Blender. However, this can be adjusted if desired.

All external models can be automatically configured and randomized in several ways via ModelCollectionRandomizer before their insertion into the world, e.g. optimization of collision geometry, estimation of (randomized) inertial properties and randomization of parameters such as geometry scale or surface friction. When processing large collections, model filtering can also be enabled based on several aspects, such as the complexity of the geometry or the existence of disconnected components. A few scripts for managing datasets can be found under scripts/utils/ directory.

This project initially investigated how 3D visual observations can be leveraged to improve end-to-end learning of manipulation skills. Octrees were selected for this purpose due to their efficiently organized structure compared to other 3D representations.

To enable the extraction of abstract features from 3D octree observations, an octree-based 3D CNN is employed. The network module that accomplishes such feature extraction is implemented in the form of OctreeCnnFeaturesExtractor (PyTorch). This features extractor is part of the OctreeCnnPolicy policy implemented for TD3, SAC and TQC algorithms. Internally, the feature extractor utilizes O-CNN implementation to benefit from hardware acceleration on NVIDIA GPUs.

Illustration of the end-to-end actor-critic network architecture with octree-based 3D CNN feature extractor.

The known limitations of this repository are listed below for your convenience.

• No parallel environments – It is currently not possible to run multiple instances of the environment simultaneously.
• Slow training – The simulation environments are computationally complex (physics, rendering, underlying low-level control, ...). This significantly impacts the ability to train agents with time and computational constraints. The performance of some of these aspects can be improved at the cost of accuracy and realism (e.g. physics_rate/step_size).
• Suboptimal hyperparameters – Although a hyperparameter optimization framework was employed for some combinations of environments and algorithms, it is a prolonged process. This problem is exacerbated by the vast quantity of hyperparameters and their general brittleness. Therefore, the default hyperparameters provided in this repository might not be optimal.
• Nondeterministic – Experiments are not fully repeatable, and even the same seed of the pseudorandom generator can lead to different results. This is caused by several aspects, such as the nondeterministic nature of network-based communication and non-determinism in the underlying deep learning frameworks and hardware.

Hardware Requirements

• CUDA GPU – CUDA-enabled GPU is required for hardware-accelerated processing of octree observations. Everything else should also be functional on the CPU.

Install Docker

First, ensure your system has a setup for using Docker with NVIDIA GPUs. You can follow install_docker_with_nvidia.bash installation script for Debian-based distributions. Alternatively, consult the NVIDIA Container Toolkit Installation Guide for other Linux distributions.

# Execute script inside a cloned repository
.docker/host/install_docker_with_nvidia.bash
# (Alternative) Execute script from URL
bash -c "$(wget -qO - https://raw.githubusercontent.com/AndrejOrsula/drl_grasping/master/.docker/host/install_docker_with_nvidia.bash)"

Clone a Prebuilt Docker Image

Prebuilt Docker images of drl_grasping can be pulled directly from Docker Hub without needing to build them locally. You can use the following command to manually pull the latest image or one of the previous tagged Releases. The average size of images is 25GB (including datasets).

docker pull andrejorsula/drl_grasping:${TAG:-latest}

(Optional) Build a New Image

It is also possible to build the Docker image locally using the included Dockerfile. To do this, build.bash script can be executed as shown below (arguments are optional). This script will always print the corresponding low-level docker build ... command for your reference.

.docker/build.bash ${TAG:-latest} ${BUILD_ARGS}

Run a Docker Container

For simplicity, please run drl_grasping Docker containers using the included run.bash script shown below (arguments are optional). It enables NVIDIA GPUs and GUI interface while automatically mounting the necessary volumes (e.g. persistent logging) and setting environment variables (e.g. synchronization of middleware communication with the host). This script will always print the corresponding low-level docker run ... command for your reference.

# Execute script inside a cloned repository
.docker/run.bash ${TAG:-latest} ${CMD}
# (Alternative) Execute script from URL
bash -c "$(wget -qO - https://raw.githubusercontent.com/AndrejOrsula/drl_grasping/master/.docker/run.bash)" -- ${TAG:-latest} ${CMD}

The network communication of drl_grasping within this Docker container is configured based on the ROS 2 ROS_DOMAIN_ID environment variable, which can be set via ROS_DOMAIN_ID={0...101} .docker/run.bash ${TAG:-latest} ${CMD}. By default (ROS_DOMAIN_ID=0), external communication is restricted and multicast is disabled. With ROS_DOMAIN_ID=42, the communication remains restricted to localhost with multicast enabled, enabling monitoring of communication outside the container but within the same system. Using ROS_DOMAIN_ID=69 will use the default network interface and multicast settings, which can enable monitoring of communication within the same LAN. All other ROS_DOMAIN_IDs share the default behaviour and can be employed to enable communication partitioning for running of multiple drl_grasping instances.

Hardware Requirements

• CUDA GPU – CUDA-enabled GPU is required for hardware-accelerated processing of octree observations. Everything else should also be functional on the CPU.

Dependencies

Ubuntu 20.04 (Focal Fossa) is the recommended OS for local installation. Other Linux distributions might work but require most dependencies to be built from the source.

These are the primary dependencies required to use this project that must be installed on your system.

• Python 3.8
• ROS 2 Galactic
• Gazebo Fortress
• Gym-Ignition
  • Please use AndrejOrsula/gym-ignition fork in order to ensure compatibility (default branch – drl_grasping).
• O-CNN
  • Please use AndrejOrsula/O-CNN fork in order to ensure compatibility (default branch – master).

All additional dependencies are either pulled via vcstool (drl_grasping.repos) or installed via pip (python_requirements.txt) and rosdep during the building process below.

Building

Clone this repository recursively and import VCS dependencies. Then install dependencies and build with colcon.

# Clone this repository into your favourite ROS 2 workspace
git clone --recursive https://github.com/AndrejOrsula/drl_grasping.git
# Install Python requirements
pip3 install -r drl_grasping/python_requirements.txt
# Import dependencies
vcs import < drl_grasping/drl_grasping.repos
# Install dependencies
IGNITION_VERSION=fortress rosdep install -y -r -i --rosdistro ${ROS_DISTRO} --from-paths .
# Build
colcon build --merge-install --symlink-install --cmake-args "-DCMAKE_BUILD_TYPE=Release"

Sourcing

Before utilizing this project via local installation, remember to source the ROS 2 workspace.

source install/local_setup.bash

This enables:

• Use of drl_grasping Python module
• Execution of binaries, scripts and examples via ros2 run drl_grasping <executable>
• Launching of setup scripts via ros2 launch drl_grasping <launch_script>
• Discoverability of shared resources

A good starting point is to simulate some episodes using random agents where actions are sampled from the defined action space. This is also useful when modifying environments because it lets you analyze the consequences of actions and resulting observations without deep learning pipelines running in the background. To get started, run the following example. It should open RViz 2 and Gazebo client instances that provide you with visual feedback.

ros2 run drl_grasping ex_random_agent.bash

After running the example script, the underlying ros2 launch drl_grasping random_agent.launch.py ... command with all arguments will always be printed for your reference (example shown below). If desired, you can launch this command directly with custom arguments.

ros2 launch drl_grasping random_agent.launch.py seed:=42 robot_model:=lunalab_summit_xl_gen env:=GraspPlanetary-Octree-Gazebo-v0 check_env:=false render:=true enable_rviz:=true log_level:=warn

You can also train your agents from scratch. To begin the training, run the following example. By default, headless mode is used during the training to reduce computational load.

ros2 run drl_grasping ex_train.bash

After running the example script, the underlying ros2 launch drl_grasping train.launch.py ... command with all arguments will always be printed for your reference (example shown below). If desired, you can launch this command directly with custom arguments.

ros2 launch drl_grasping train.launch.py seed:=42 robot_model:=panda env:=Grasp-OctreeWithColor-Gazebo-v0 algo:=tqc log_folder:=/root/drl_grasping_training/train/Grasp-OctreeWithColor-Gazebo-v0/logs tensorboard_log:=/root/drl_grasping_training/train/Grasp-OctreeWithColor-Gazebo-v0/tensorboard_logs save_freq:=10000 save_replay_buffer:=true log_interval:=-1 eval_freq:=10000 eval_episodes:=20 enable_rviz:=false log_level:=fatal

Remote Visualization

To visualize the agent while training, separate RViz 2 and Gazebo client instances can be opened. For the Docker setup, these commands can be executed in a new drl_grasping container with the same ROS_DOMAIN_ID.

# RViz 2 (Note: Visualization of robot model will not be loaded using this approach)
rviz2 -d $(ros2 pkg prefix --share drl_grasping)/rviz/drl_grasping.rviz
# Gazebo client
ign gazebo -g

TensorBoard

TensorBoard logs will be generated during training in a directory specified by the tensorboard_log:=${TENSORBOARD_LOG} argument. You can open them in your web browser using the following command.

tensorboard --logdir ${TENSORBOARD_LOG}

(Experimental) Train with Dreamer V2

You can also try to train some agents using the model-based Dreamer V2 algorithm. To begin the training, run the following example. By default, headless mode is used during the training to reduce computational load.

ros2 run drl_grasping ex_train_dreamerv2.bash

After running the example script, the underlying ros2 launch drl_grasping train_dreamerv2.launch.py ... command with all arguments will always be printed for your reference (example shown below). If desired, you can launch this command directly with custom arguments.

ros2 launch drl_grasping train_dreamerv2.launch.py seed:=42 robot_model:=lunalab_summit_xl_gen env:=GraspPlanetary-ColorImage-Gazebo-v0 log_folder:=/root/drl_grasping_training/train/GraspPlanetary-ColorImage-Gazebo-v0/logs eval_freq:=10000 enable_rviz:=false log_level:=fatal

Once you train your agents, you can evaluate them. Start by looking at ex_evaluate.bash, which can be modified to fit your trained agent. It should open RViz 2 and Gazebo client instances that provide you with visual feedback, while the agent's performance will be logged and printed to STDOUT.

ros2 run drl_grasping ex_evaluate.bash

After running the example script, the underlying ros2 launch drl_grasping evaluate.launch.py ... command with all arguments will always be printed for your reference (example shown below). If desired, you can launch this command directly with custom arguments. For example, you can select a specific checkpoint with the load_checkpoint:=${LOAD_CHECKPOINT} argument instead of running the final model.

ros2 launch drl_grasping evaluate.launch.py seed:=77 robot_model:=panda env:=Grasp-Octree-Gazebo-v0 algo:=tqc log_folder:=/root/drl_grasping_training/train/Grasp-Octree-Gazebo-v0/logs reward_log:=/root/drl_grasping_training/evaluate/Grasp-Octree-Gazebo-v0 stochastic:=false n_episodes:=200 load_best:=false enable_rviz:=true log_level:=warn

The default hyperparameters for training agents with TD3, SAC and TQC can be found under the hyperparams directory. Optuna can be employed to autotune some of these parameters. To get started, run the following example. By default, headless mode is used during hyperparameter optimization to reduce computational load.

ros2 run drl_grasping ex_optimize.bash

After running the example script, the underlying ros2 launch drl_grasping train.launch.py ... command with all arguments will always be printed for your reference (example shown below). If desired, you can launch this command directly with custom arguments.

ros2 launch drl_grasping optimize.launch.py seed:=69 robot_model:=panda env:=Grasp-Octree-Gazebo-v0 algo:=tqc log_folder:=/root/drl_grasping_training/optimize/Grasp-Octree-Gazebo-v0/logs tensorboard_log:=/root/drl_grasping_training/optimize/Grasp-Octree-Gazebo-v0/tensorboard_logs n_timesteps:=1000000 sampler:=tpe pruner:=median n_trials:=20 n_startup_trials:=5 n_evaluations:=4 eval_episodes:=20 log_interval:=-1 enable_rviz:=true log_level:=fatal