The ACRV Picking Benchmark (APB)

Baseline System

Hardware  Software  Benchmark 

Baseline System Abstract

We present a benchmark shelf picking system primarily composed of inexpensive, readily available components. The platform leverages a single seven degree of freedom arm of a static Baxter robot. To promote comparison against a variety of systems (in the ACRV Picking Benchmark), a single-arm setup was chosen.

Download:   the paper,   the baseline code, and end-effector model

Hardware Details

The baseline system extends a Baxter Research Robot (from Rethink Robotics) with additional perception capabilities, a custom end-effector with two suction cups at right angles, a small kinetic vacuum pump, and an Intel NUC PC mounted on the robot’s elbow.

End Edffector Design

The robot’s custom-designed gripper provides a mounting point for two suction cups, two vacuum lines and a small RGB-D camera. The design is available under a Creative Commons BY-SA 3.0 license: CAD models
download end-effector schmeatics
The gripper consists of two 3D-printed components – one connected to the wrist of the robot, the second to house the two suction cups – connected by two parallel PVC pipes of 27cm length (diameter of 20mm). Existing gripper mount points allowed for our gripper to be attached to Baxter non-destructively.
Suction cups provide a simple and effective way to grasp a wide variety of objects. Our design includes two separate cups to increase the number of reachable grasp points. Considering an object-filled shelf, the additional right-angled suction cup enables object grasps from above, as well as, side ways picks of objects leaning against the wall, especially books and DVDs. Aligning a suction cup with the axis of rotation of the wrist creates the capability to grasp objects from the side.




Sensor Systems

First the robot localises the shelf by using the Kinect2 mounted on the top of its head. This provides a global position of the robot wrt. the shelf.
An Intel RealSense SR300 (RGB-D camera), mounted at a fixed-angle location, provides local shelf sensing. The angle chosen was empir- ically found by maximising closeness to the front face of the shelf and reducing the joint-angle distance required to reposition between look-into-shelf and pre-grasp pose.







Software Details

We leverage the Robot Operating System (ROS) framework and additional open source software to speed up development of the system and promote modularity, standardisation and reproducibility. The baseline system’s source code is publicly available at: GitHub.
Software packages integrated in the system include: MoveIt! and OMPL (Motion Planning), TRAC-IK (Inverse Kinematics Solver), Point Cloud Library, iai_kinect2 driver, librealsense driver, SMACH (State Machine), and many more.




Perception Pipeline

First the robot localises the shelf by using the Kinect2. The robot then chooses the next object to pick from the work order provided (in a JSON file) and moves the robot’s end effector to a pre-recorded scan pose. During a diamond- shaped scanning operation (parallel to the shelf’s front), point clouds, provided by an Intel RealSense on the wrist, are recorded. A fused cloud is then sent into the perception pipeline, which segments and identifies the target object.
grasping pipeline



Grasping Pipeline

The objective of the grasping pipeline is to provide reach- able grasp points on the object to pick. Provided a point cloud segment, candidate grasp points are generated by smoothing the cloud, estimating it’s boundaries and computing point normals in a grid-like pattern across the surface. This step is fast to perform and includes hyper-parameters for the suction cup dimensions to reduce/increase the number of candidate points generated. A grasp selection process follows. First, inverse kinematics for each candidate are checked at both pre-grasp (5 cm above) and grasp poses to ensure the robot can reach the object along the candidate point normal. Reachable grasps are then ranked with heuristics such as distance to the clouds boundary, curvature at the normal and distance to the walls of the bin. These heuristics were chosen to give our system the best chance of grasping an object in its current pose.