Uncertainty-Aware Manipulation Planning accepted to RA-L and IROS 2022
A paper of the OMRON SINIC X Robotics group was accepted for publication in IEEE Robotics and Automation Letters. The source code is published here, along with a robot manipulation setup in ROS (link).
- Felix von Drigalski, Kazumi Kasaura, Cristian C. Beltran-Hernandez,
Masashi Hamaya, Kazutoshi Tanaka and Takamitsu Matsubara, “Uncertainty-aware manipulation planning using gravity and environment geometry” (see full paper)
This paper is the capstone of a research project for reliable robot manipulation, motivated not least by the World Robot Summit Assembly Challenge. In robotic assembly, small positioning errors can cause task failure. After grasping an object, its position usually has errors from e.g. a noisy detection or sensor input.
Jigs are often used to position the part after grasping. However, jigs are expensive and space-inefficient, limiting the flexibility of the robot system. We propose a method to determine the position of a grasped object to high sub-millimeter precision without expensive sensors. Our method can be applied to rigid objects and requires only a mesh representation.
In this paper, we present three new actions using the environment geometry (Push, Place and Grasp), as well as a planner to determine the best course of action.
How it works
After grasping, we let the robot perform actions that constrain the possible positions of the objects in the gripper.
We represent the object’s position as a 6D pose and a normal distribution modelling the pose uncertainty. Each action decreases the uncertainty about the object’s pose.
Finding the right sequence of actions
Now that the effect of each action can be calculated, we can find a sequence of actions that minimizes the uncertainty.
At the end of the sequence, the object’s position in the gripper is known with high precision. The new Push, Place and Grasp actions allow for more efficient action sequences that eliminate the uncertainty entirely.
In our experiments, we achieved sub-mm accuracy with these action sequences in over 95% of cases, and accuracy within 3 mm over 99% of the time.
Using this code in your work
We have released the source code for this project as a ROS repository here. It can be used stand-alone without the ROS wrapper.
The ROS wrapper allows the code to be used with the system released here by Team O2AC after the World Robot Summit Assembly Challenge. The repository offers a two-armed robot manipulation setup which can reproduce our experiments in simulation.
Limitations
We make some assumptions when evaluating the actions’ outcomes:
- Objects don’t roll away once placed
- Objects don’t topple during grasping (friction is ignored)
- Push, Place and Grasp surfaces are planar
These assumptions are more easily fulfilled than it appears — rolling can be prevented easily by a thin layer of foam and toppling did not occur in our experiments. For applications where it is necessary, the permissible actions can be constrained to ensure that they operate within the expected regime.
Future work
Our investigation leaves us with open questions:
- Can a grasp be guaranteed to succeed given an object pose and uncertainty? If not, can the likelihood of success be estimated efficiently?
- Can the approach be used for grippers with more fingers, non-planar fingers or suction cups/pads?
Do these questions pique your interest, and you are a student looking for an exciting project in Tokyo? Then you should apply for an internship!
