12 Open Problems in Industrial Robotics
Although there are countless open problems in robotics, we hope that
by defining what we consider to be the 12 most important open
problems in RISC
robotics, effort can be concentrated in those areas. In the past
year, a number of new problems have emerged. Our objective is to
identify new theoretical problems with near-term relevance for
industrial robotics.
We welcome feedback and/or comments to this list. In addition we would
like to add pointers to people currently working on particular
problems so that open communication is maintained and duplicate work
does not occur. Also, pointers to relevant papers should be added. We
welcome submissions of work in progress in this location for feedback.
For any comments or changes please contact Eric
Paulos.
The current list came out of a workshop held at Adept by Ken Goldberg in December
1994 and as he points out, "The goal is to specify problems as crisply
as possible; the list below is just a first cut." In addition, full
credit for this exceptional list should go to Ken Goldberg.
The list below of the 12 open problems is presented in no particular
order. The challenge for each problem is is to define optimality and
then find an efficient algorithm or prove that no such algorithm exists.
[1] Computing Line and Ray Hull
Description coming soon....
[2] RISC Sensor Design
Given a family of parts, find a geometric arrangement of cross-beam,
parallel-beam, or a small set of sparse-depth sensors to optimally
determine pose or to distinguish between parts in this family. Model
Generation for RISC Sensors. Given a polygonal part profile,
efficiently generate a lookup table for estimating pose using a
parallel-beam sensor.
[3] Design of Gripper Jaws
Given a set of grasps on a known part or set of parts, design a shape
for each gripper jaw that makes all necessary contacts without
collisions and optimizes grasp stability for each grasp.
Work in progress by Eric Paulos and John Canny
[4] Autoplace
Given a desired trajectory for the robot wrist and a model of robot
kinematics/dynamics, locate the robot base to minimize total travel
time (an analytic measure might be weighted travel for all joints).
Autoplace.1. As above, but consider the base fixed and locate targets
subject to workspace constraints. Autoplace.2 Given desired positions
for parts in workspace, locate the camera to minimize parallax
effects, obstructions, etc. while permitting maximal resolution.
Autoplace.3 As above, but let all positions vary simultaneously.
[5] Pose Estimation for Tipped Parts
Known parts arrive in pallets but may be tipped 20-30 degrees from a
nominal pose. Similar effects arise due to camera parallax as pallet
height changes (eg, when unstacking layers from a carton). This is
far more constrained than general 3D vision. Find a low-cost, fast
(<100msec), and reliable method for localizing and grasping such
parts.
[6] Predicting Feeder Throughput
To predict throuput for a flexible part feeder based on dropped parts,
we need to predict the distribution of stable poses. Given a CAD part
model with center of gravity, find the set of stable poses of the part
and estimate a random distribution on this set. A quasi-static model
projects cog onto a unit sphere and compares projected area of each
face of the part's convex hull, taking care to propagate probability
from unstable faces onto stable faces. Improve this with a dynamic
model that incorporates friction and coefficient of restitution but
runs in polynomial time.
[7] Find-Space for Dropped Parts
To simulate a flexible part feeder, we model parts dropping from a
known height onto a worktable. The current model proposes a random
final position for the part depending on drop height and coefficient
of restitution, but uses generate-and-test to avoid collisions with
existing parts. Given a collection of obstacles, find a pose where a
new part can be placed without collision. Is there an efficient
on-line algorithm?
[8] Grasp Planning for Parallel-Jaw Gripper
For each stable pose of a given polyhedral part, find a grasp assuming
two point contacts. The grasp should minimize reliance on friction
and should be robust to small errors in part pose, for example using
Brost-like push-grasping.
[9] Grasp Planning for Pivoting Gripper
As above but consider a pivoting axis between the contact points.
Given initial and desired final pose of the part, find a grasp that
will permit the part to be pivoted into the desired final pose.
Rao/Kriegman/Goldberg give an O(n\logn) algorithm (for polyhedral
parts with n faces) for finding all pivot grasps that rely on gravity
to cause pivoting.
[10] On-Line Motion/Grasp-Circuit Planning
Given parts arriving at unpredictable times on a conveyor belt, give
an algorithm for rapidly scheduling efficient grasp paths. Gripper
may have multiple suction cups that must be "loaded" during each pass:
Given a robot gripper with k suction cups and a collection of n parts
on a conveyor belt in poses determined by an overhead vision system,
rapidly approximate an optimal path for loading the suction cups and
depositing the collection in a pallet. A variant considers possible
collisions with a second arm. Algorithm must trade off plan time
vs. execution time.
[11] DFX
Design for X, where X is a property related to manufacturing such as
feedability or fixturability. Quantify such a property and then
develop an algorithm that takes an existing geometric design as input
and generates a new design that gives better performance for X.
[12] 3D Fixture Design
Given part geometry, design a minimal set of modular components
capable of holding a broad class of 3D parts. Then develop an
algorithm that, given part geometry, will find an optimal arrangement
of these components.
UCB Robotics Lab / Eric Paulos /
26 May 1995