Deep Grasp Library for ROS2

The primary philosophy behind this repo is that a full, deployable, and usable AI solution extends beyond the AI model itself. dgl_ros is a light-weight C++ library for researchers to package their deep grasp models into a ROS node, ready to be run and experimented with by the robotics community. dgl_ros_models contains examples of popular deep grasp models such as GPD and Contact Graspnet packed and ready to use.

Overview

dgl_ros addresses the daunting heterogeneity in deep grasp model inputs, action spaces, and library dependencies through abstraction and composition. Specifically, we package a minimum solution as an agent that composes an observer service and an actor service. This is hardly enough, however, for any realistic use-case which requires an additional supervisor service to provide a label or reward signal for fine-tuning.

Features

• Base class with two virtual methods, ObsFromSrcs and ActionFromObserver, to create a Ros2 action server for an arbitrary deep grasp model
• End-to-end working inference nodes for GPD and Contact Graspnet
• No configuration needed to run the provided deep grasp action servers outside of a yaml file.
• MoveIt Task Constructor stage to interface with the action server from a higher-level planning pipeline

Coming Soon

1. Inference node for pretrained Google's RT-X (Robot Transformers)
2. Supervisor service to label samples with reward or success/failure
3. DatasetGenerator class to create a dataset for offline learning in pytorch
4. Training module to collect and label data from parallel simulations
5. Python bindings

Demos in Simulation and Real Life

Requirements

This has been tested with the following dependencies

1. Ubuntu 20.04
2. ROS2 Humble
3. C++ 17

Example Usage

Contact Graspnet

1. Install cgn-pytorch: pip install cgn-pytorch
2. Install pybind11 sudo apt -y install python-pybind11
3. Clone this library into your $ROS_WS/src directory: git clone https://github.com/sebbyjp/dgl_ros.git
4. Install dependencies: rosdep install --from-paths src --ignore-src --rosdistro ${ROS_DISTRO} -y
5. Build this library from your $ROS_WS directory: colcon build --symlink-install --base-paths src/dgl_ros/
6. In a terminal, run ros2 run dgl_ros_models contact_graspnet --ros-args -p src_topic0:=$ROS_TOPIC_GENERATING_POINTCLOUDS -p cgn_config_path:=$PATH_TO_YOUR_CONFIG

GPD

1. Follow instructions to install GPD
2. Clone this library into your $ROS_WS/src directory: git clone https://github.com/sebbyjp/dgl_ros.git
3. Install dependencies: rosdep install --from-paths src --ignore-src --rosdistro ${ROS_DISTRO} -y
4. Build this library from your $ROS_WS directory: colcon build --symlink-install --base-paths src/dgl_ros/
5. In a terminal, run ros2 run dgl_ros_models gpd --ros-args -p src_topic0:=$ROS_TOPIC_GENERATING_POINTCLOUDS -p gpd_config_path:=$PATH_TO_YOUR_GPD_CONFIG

Note: Make sure you set use_sim_time:=true if you are running a simulation. For this example we send a goal from another terminal so it should be omitted.

6. In a separate terminal, run ros2 action send_goal /sample_grasp_poses dgl_ros_interfaces/action/SampleGraspPoses "{action_name: 'sample_grasp_poses'}"

Parameters for the action server node:

• src0_topic, src0_frame, ... srcN_topic, srcN_frame
• world_frame
• TODO: List in more detail

Concepts

Classes to Represent Data

Src: A ROS message typically published from simulation or ROS sensor control plugins (e.g. sensor_msgs/msg/Image, sensor_msgs/msg/Pointcloud2)

Observation: A ROS message produced from one or more Src messages that has all the data needed to input into a deep learning model.

Classes to Interact with the Environment

The following classes make up the environment interface in RL libraries such asOpenAI's gym and Google's TF-Agents. They are the bridge between the real world and the learning model.

Observer: A ROS node that subscribes to one or more Src topics, creates an Observation when requested, and stores a cache of Observations which can be accessed by id.

Actor: A ROS action server that requests Observations from an Observer and produces an action (e.g. a list of grasp poses) . This class wraps a deep learning model (representing the policy in RL) and converts the into a ROS message. It is accessed via the ROS action server interface.

Supervisor: A ROS server that produces a label or reward signal for an Observation and ROS action. It is called via the ROS service interface or directly through member functions.

Design Decisions

• To integrate a new deep learning model, minimum extra configuration should be required:

  • Define a new ROS action (e.g. ros_dgl_interfaces/action/SampleGraspPoses.action)
  • Define a new ROS message representing an Observation
  • Extend the Agent class and implement the ObsFromSrcs() and ActionFromObs() methods

• The primary interface for a Deep Grasp Agent is ROS2 topics and services. That way inference can be run by any compatible ROS2 project.

• Class structure should follow Reinforcement Learning concepts used by OpenAI's gym and Google's TF-Agents. By committing to the RL framework, data is represented as flexibly as possible and includes open-loop models and datasets like Contact Graspnet.