IEEE Robotics & Automation Magazine - September 2015 - 49

given at http://rll.eecs.berkeley.edu/ycb/. A snapshot of this
setting can be seen in Figure 14.
YCB Block Pick-and-Place Protocol and Benchmark
Manual dexterity and the manipulation of small objects are
critical skills for robots in several contexts. The block pickand-place protocol is designed to test a robot's ability to
grasp small objects and transfer them to a specified location.
This task is an important test of both arm and gripper hardware and motion planning software, as both contribute to
overall dexterity. Points are awarded based on completion
and precision of the manipulation. We executed this test on
the HERB robot [61], as seen in Figure 15. An image of the
printed layout with the placed blocks after task completion
can be seen in Figure 16. The results show that the robot is
not able to succeed in precise pick-and-place task. The main
reason is the utilized open-loop grasping approach. The
robot executes a robust push grasp strategy, which allows it
to grasp the blocks successfully. However, the pose of the
block with respect to the gripper is not known precisely
after the grasp. This prevents placing the blocks accurately
to the target locations.
YCB Peg-Insertion Learning-Assessment
Protocol and Benchmark
The peg-insertion learning-assessment benchmark is designed to allow comparison among various learning tech-

niques. The benchmark measures the performance of a
learned peg-insertion action under various positioning perturbations. The perturbations are applied by moving the peg
board to a random direction for certain amount of distance.
We applied this benchmark to assess the performance of a
learned linear-Gaussian controller using a PR2 robot [66]
(Figure 17). The state of the controller consists of the joint
angles and angular velocities of the robot as well as the positions and velocities of
three points in the space
The Model T42 is able
of the end effector (three
points to fully define a
to provide stable power
rigid body configuration).
No information is availgrasps for large objects
able to the controller at
run time except for this
and precision grasps
state information. The results show that the
for small objects.
learned controller shows
reasonable performance,
with four successes out of
ten trials, for the case of 5-mm position perturbation to a
random direction. This success rate can be achieved by executing the controller for only 1 s. However, the performance
does not improve, even if the controller is run for a longer
period of time. In the case of 10-mm position perturbation,
the controller fails completely. We are planning to learn the
same task with different learning techniques and compare
their performances using the benchmark.

(a)

(b)

Figure 14. The simulation environment for the table-setting
benchmark. This environment can be spawned using the URDF
provided at http://rll.eecs.berkeley.edu/ycb.

Figure 16. (a) and (b) The results of the block pick-and-place
benchmark.

Figure 15. The HERB robot implementing the peg-insertion
learning-assessment benchmark.

Figure 17. The PR2 executing the peg-insertion learningassessment benchmark.

September 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

http://rll.eecs.berkeley.edu/ycb/ http://rll.eecs.berkeley.edu/ycb

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2015