About Two Motion Planning Papers

36
Terry Taewoong Um ([email protected]) University of Waterloo Department of Electrical & Computer Engineering Terry Taewoong Um MOTION PLANNING 1

Transcript of About Two Motion Planning Papers

PowerPoint

Terry Taewoong Um ([email protected]) University of Waterloo Department of Electrical & Computer Engineering

Terry Taewoong Um

Motion Planning1

1

Terry Taewoong Um ([email protected])My previous research2

2008-2010 RRT planning2011-2012 Exoskeleton

2

Terry Taewoong Um ([email protected])ContentsPreliminaries

Dynamically-Stable Motion Planning for Humanoid Robots (2002)

Whole-body Motion Planning with Centroidal Dynamics and Full Kinematics (2014)3

3

Terry Taewoong Um ([email protected])Contents4Preliminaries

4

Terry Taewoong Um ([email protected])5Why do we need planning?What would be most important in planning? Optimality Stability Completeness

5

Terry Taewoong Um ([email protected])6CompletenessIf there are solutions for a problem, a planning algorithm can find a solution. If there is no solution for a problem, a planning algorithm return failure. We can say This planning algorithm is complete

Probabilistic completeness : With enough samples, the probability that it finds an existing solution converges to one.

Most sampling-based motion planning algorithm aims at guaranteeing probabilistic completeness (One of the easiest way is using randomness)

6

Terry Taewoong Um ([email protected])7Handling constraints

Nonholonomic constaintsConstrained manifold

7

Terry Taewoong Um ([email protected])8Tangent space RRT (TS-RRT)

8

Terry Taewoong Um ([email protected])9Questions: Constraints

Q) Would it be beneficial for the planning software to compute more than just the first solution it finds?

Q) How do we make sure there is no conflict between constraints or having conflict is not important here?

Q) Perhaps some constraints are more important than others. Is there any benefit to assign priority to the constraints here? If yes, how can we implement a hierarchy for constraints (like the one we learned in full body control paper)?

Q) What is the secret that the algorithm worked well in such hard tasks? Say for example in monkey bar task, any small deviation from desired posture or torque will lead to failure.

9

Terry Taewoong Um ([email protected])Contents102. Dynamically-Stable Motion Planning for Humanoid Robots (2002)

10

Terry Taewoong Um ([email protected])11OverviewFirst phase : statistically-stable, collision-free path

Second phase : smooths and transforms this path into dynamically-stable trajectory

Given - Environment Model (collision checking is possible) - Initial / Goal posture (collision-free & statistically-stable) - Support base : The location of the supporting foot

Robot model :

Q) Why do we needbothP and Crepresentations?Q) How would this be used with support foot movement? Can you switch feet or take steps?

11

Terry Taewoong Um ([email protected])12Problem FormulationQuery & Solution :

Balance & Torque constraints : q is statistically-stable ifThe projection of X(q) along the gravity vector g lie within the area of support SP (the convex hull of all points of contact)The joint torques do not exceed the maximum torque bound

Q) In what situations do we need the second condition for static-stability? Obstacle =

Dynamic stability : statically-stable slowing down dynamically-stable AutoBalancer online balance compensation scheme

Q) How can any statically-stable trajectory betransformed into a dynamically-stable trajectory by arbitrarily slowing down the motion?

12

Terry Taewoong Um ([email protected])13Path searchThere is no method for explicitly representingBut we have collision checker & stability checker

Distance metric

- Higher weights to the links with greater mass and proximity to the trunk

Q) What does that mean how mass andproximityto torso relates to similarity in configurations? Q) Why in equation (2) did they assign a higher value to joints closer to the torso? What effect does this have on the motion? Q) Do you think you can come up with better distance metric using lie group?

13

Terry Taewoong Um ([email protected])14RRT

Q) Is it possible for RRT to enter localminimasand get stuck? Is this when the failure after a preset time limit happens?

14

Terry Taewoong Um ([email protected])15RRT for stable motions

- Difference : instead of

- Reached / Trapped / Advanced- The tree attempts to make an incremental motion toward q If , add q to the tree

Q) Why do we have to pick q among statically stable postures despite q will not be part of path unless(q,qnear) # of possible sequences

Complementarity Constraints

normalcontact Fdist. from the contact pt

When contact forces exist at a certain contact point and the contact torque is always zeroif the normal contact force exists, i.e. , the body is in contact, then the tangential displacement of the body between the two adjacent knots must to be zero;

if the body moves in the tangential directions, then the normal force has to be zero, i.e., the body is not in contact

30

Terry Taewoong Um ([email protected])31unscheduled contact sequenceThey first put L2 norm then removed it from cost function. Why in the first place they introduced it?

Q) In case of unscheduled contact sequence, what the solver does is to search over all possible contact sequences. Does that mean we should define all possible contact sequences for the algorithm to search among them?

Q) How do defining complementary constraints help to avoid predefining of sequence modes? New objective function

With the extra complementarity constraints and the revised objective function, the solver can search over all possible contact sequences

31

Terry Taewoong Um ([email protected])32Experiment

32

Terry Taewoong Um ([email protected])33Experiment

Jumping (used the fixed contact sequence)Q) Why in jumping off a box we need to put kinematic constraint that feet avoid collision box? Didnt we already take care of collision in equation 5?

Running Q) When they are discussing running they say they plan a half stride and mirror it. But if this was applied to obstructed environments wouldn't this not be valid as your strides would need to change to avoid obstacles? - stride length / speed : 2m , 2m/s

33

Terry Taewoong Um ([email protected])34ExperimentMonkey bars (w/ and w/o a fixed mode sequence)Salmon ladder

Little dogQ) Is there any planning algorithm that works for humanoid robots but is not general? I assume that due to complexity of humanoids compared to other robots any planning algorithm for humanoids can be applied to other robots by some modifications. If you agree why generality is highlighted in these papers? Q) Regarding the validation, how strong do you think is to evaluate the proposed method only by referring to robot snapshots and videos and having not enough numerical results?

34

Terry Taewoong Um ([email protected])35QuestionsQ) In what scenarios would this method not work well?

Q) Is this method only implemented in simulation?

Q) Do you think this method will work on an actual atlas robot or is there a reason they only did this in simulation?

Q) What would be the main challenges in implementing the proposed method in real ATLAS robot? Do you think itwillworkat all?

Q) What does it mean that the C++ code is about two orders of magnitude faster than the MATLAB code for NLPs? Does it mean that Matlab is less capable to deal with NLPs or is that their coding style results in faster computation in C++?

35

Terry Taewoong Um ([email protected])36Thank youContents

36