KOMO-3: Manipulation Modelling & Execution

The discussed components (KOMO, BotOp, NLP_Solver, RRT) provide basic ingredients for manipulation planning and execution. This tutorial is about how to practically use these in typical manipulation settings.

The first focus is on manipulation modelling. While KOMO provides a very powerful abstract framework to define all kinds of constraints, here we discuss what are concrete useful constraints for typical actions, e.g., picking and placing a box, or capsule. The ManipulationModelling class is meant to translate between typical actions and the abstract KOMO specification of the corresponding constraints.

The second focus is on the whole pipeline. We follow a basic sense-plan-act pipeline (not yet a fully integrated reactive framework such as SecMPC). To be more detailed, we assume the following basic steps in each loop: * Perception: Update the configuration model to be in sync with the real world - using perception. * Discrete decisions (task planning): Decide on discrete next actions, such as which object to pick or place next. * Waypoint planning: Model the manipulation constraints for the next few actions and solve them to get a plan for the next few waypoints. * Path planning: Create a fine-grained path/trajectory between waypoints, sometimes justing quick interpolation & optimization, sometimes using full fledge path finding (bi-directional RRT). * Execution: Sending the path to BotOp for running it on the real system.

We neglect perception and discrete decision making here.

Manipulation Modelling

We start with discussing manipulation modelling for standard box/cylinder grasping and placing.

import robotic as ry
import numpy as np
import random
import time

# this import the local manipulation.py .. to be integrated in robotic..
import manipulation as manip

A basic configuration with a box and cylinder:

C = ry.Config()
C.addFile(ry.raiPath('scenarios/pandaSingle.g'))

C.addFrame('box') \
    .setPosition([-.25,.1,1.]) \
    .setShape(ry.ST.ssBox, size=[.06,.06,.06,.005]) \
    .setColor([1,.5,0]) \
    .setContact(1)

C.addFrame("capsule") \
    .setShape(ry.ST.capsule, [.2,.02]) \
    .setPosition([.25,.1,1.]) \
    .setColor([1,.5,0]) \
    .setContact(1)

# for convenience, a few definitions:
qHome = C.getJointState()
gripper = "l_gripper";
palm = "l_palm";
box = "box";
table = "table";
boxSize = C.frame(box).getSize()

C.view()

Look into the definition of ManipulationModelling class! You see that this class provides wrapper methods to setup a komo problem. The following demonstrate the methods provided to model box and cylinder grasping IK problems:

Box centered top grasp

There are 6 possible orientation of an orthonormal centered box grasp. Have a look at the grasp_top_box method!

C.setJointState(qHome)
for orientation in ['xy', 'xz', 'yx', 'yz', 'zx', 'zy']: #loops over the 6 possible grasps
    # setup the manipulation problem
    man = manip.ManipulationModelling(C)
    man.setup_inverse_kinematics()
    man.grasp_top_box(1., gripper, box, orientation)

    # solve it
    pose = man.solve()

    # check feasibility and display
    if man.ret.feasible:
        C.setJointState(pose[0])
        C.view(True, f'grasp_top_box with orientation {orientation}\nret: {man.ret}')
    else:
        print(' -- infeasible')

Box general grasp

We do not have to grasp a box in the center or orthonormally. We can only specify along which axis the fingers should press, and that they need to be inside a margin of the box sides. Have a look at the grasp_box method! To illustrate the gained degrees of freedom, we also impose a random bias (leading to different solutions in nullspace):

C.setJointState(qHome)
limits = C.getJointLimits()
for orientation in ['x', 'y', 'z']:
    for i in range(10):
        # setup the manipulation problem
        man = manip.ManipulationModelling(C)
        man.setup_inverse_kinematics()
        # ... with random bias in joint space
        qBias = limits[0]+np.random.uniform(qHome.shape)%(limits[1]-limits[0])
        man.bias(1., qBias, 1e0)
        # ... and general, non-centered box grasping
        man.grasp_box(1., gripper, box, palm, orientation, margin=.02)

        # solve
        pose = man.solve()

        # if feasible, display
        if man.ret.feasible:
            C.setJointState(pose[0])
            C.view(True, f'grasp_box with orientation {orientation}\nret: {man.ret}')
        else:
            print('-- infeasible', i, orientation)

Cylinder grasp

A cylinder (or capsule) can be grasped by ensuring the finger axis is normal to the cylinder’s axis – have a look at the grasp_cylinder method. Again, a demo with random bias to show the variety of grasps modelled that way:

C.setJointState(qHome)
limits = C.getJointLimits()
for i in range(10):
    # setup the manipulation problem
    man = manip.ManipulationModelling(C)
    man.setup_inverse_kinematics()
    qBias = limits[0]+np.random.uniform(qHome.shape)%(limits[1]-limits[0])
    man.bias(1., qBias, 1e0)
    man.grasp_cylinder(1., gripper, 'capsule', palm)

    # solve
    pose = man.solve()

    # if feasible, display
    if man.ret.feasible:
        C.setJointState(pose[0])
        C.view(True, f'grasp_cylinder\nret: {man.ret}')
    else:
        print('-- infeasible', i, orientation)

Sequential Manipulation Modelling

Sequential manipulation modelling is special, as in some phases the manipulated objects move with the manipulator. Internally, komo models this with a mode switch (where an object becomes attached to the manipulator with a stable (but optimizable) relative transform).

Using the ManipulationModelling class, the setup_pick_and_place_waypoints method creates a two-time-steps komo problem where the relative object-gripper position is constrained to be the same in the 1st and 2st step (as it is parameterized by a shared relative kinematic joint). The grasp_box method ensures that the solution also fulfils grasp constraints in the first time step; and the place_box method ensures that the solution also fulfils placement constraints in the second time step. The additional target_relative_xy_position is optional, so see placement to explicit xy-positions on the table. Have a look at the definitions of all these methods.

C.setJointState(qHome)

for i in range(10):
    grasp_ori = random.choice(['x', 'y', 'z'])
    place_ori = 'z' #random.choice(['x', 'y', 'z'])
    info = f'pnp {i}, grasp orientation {grasp_ori}, place orientation {place_ori}'
    print('===', info)

    # setup manipulation problem
    man = manip.ManipulationModelling(C)
    man.setup_pick_and_place_waypoints(gripper, box)
    man.grasp_box(1., gripper, box, palm, grasp_ori)
    man.place_box(2., box, table, palm, place_ori)
    man.target_relative_xy_position(2., box, table, [(i%5)*.1-.2, .2])

    # solve
    q = man.solve()

    # if feasible, display (including 'fake' simulation with kinematic attach)
    if man.ret.feasible:
        C.setJointState(q[0])
        C.view(True, f'{info}\nwaypoint 0\nret: {man.ret}')
        C.attach(gripper, box)
        C.setJointState(q[1])
        C.view(True, f'{info}\nwaypoint 1\nret: {man.ret}')
        C.attach(table, box)
        C.setJointState(qHome)
        C.view(True, 'back home')
    else:
        print(' -- infeasible')

del man
C.frame('box').setPosition([-.25,.1,1.])
C.view()

Path generation

Once solutions to the manipulation keyframes/waypoints are available, the next step is to generate motion between them. We can use sample-based path finding (bi-directional RRT) and/or smooth motion optimization for this.

Smooth point-to-point motion

The following demonstrates smooth point-to-point motion between box grasps, there the motion is additionally constrains the endeffector to retract and approach:

C.setJointState(qHome)
limits = C.getJointLimits()
verbose = 0

for i in range(20):
    qStart = C.getJointState()

    # choose a random grasp orientation
    orientation = random.choice(['x', 'y', 'z'])
    print('===', i, 'orientation', orientation)

    # setup the grasp problem
    man = manip.ManipulationModelling(C)
    man.setup_inverse_kinematics(accumulated_collisions=True)
    man.grasp_box(1., gripper, box, palm, orientation)

    # solve
    pose = man.solve()
    print('    IK:', man.ret)

    # if feasible, display; otherwise try another grasp
    if man.ret.feasible:
        if verbose>0:
            C.setJointState(pose[0])
            C.view(True, f'grasp {i} with orientation {orientation}\nret: {man.ret}')
    else:
        print('  -- infeasible')
        C.setJointState(qStart)
        continue

    # setup the motion problem
    man = manip.ManipulationModelling(C, helpers=[gripper])
    man.setup_point_to_point_motion(qStart, pose[0])
    man.retract([.0, .2], gripper)
    man.approach([.8, 1.], gripper)

    # solve
    path = man.solve()
    print('  path:', man.ret)

    # if feasible, display trivially (no real execution in BotOp here)
    if man.ret.feasible:
        for t in range(path.shape[0]):
            C.setJointState(path[t])
            C.view(False, f'grasp {i} with orientation {orientation}, path step {t}\n{man.ret}')
            time.sleep(.05)
        C.view(verbose>0, f'path done')
    else:
        print('  -- infeasible')

del man

Integrated Example

Let’s start with an integrated example, where the robot endlessly loops through picking and placing a box on a table.

import robotic as ry
import manipulation as manip
import numpy as np
#from importlib import reload
import time
import random

We define a basic configuration with box on the table:

C = ry.Config()
C.addFile(ry.raiPath('../rai-robotModels/scenarios/pandaSingle.g'))

C.addFrame("box", "table") \
    .setJoint(ry.JT.rigid) \
    .setShape(ry.ST.ssBox, [.15,.06,.06,.005]) \
    .setRelativePosition([-.0,.3-.055,.095]) \
    .setContact(1) \
    .setMass(.1)

C.addFrame("obstacle", "table") \
    .setShape(ry.ST.ssBox, [.06,.15,.06,.005]) \
    .setColor([.1]) \
    .setRelativePosition([-.15,.3-.055,.095]) \
    .setContact(1)

C.delFrame("panda_collCameraWrist")

# for convenience, a few definitions:
qHome = C.getJointState()
gripper = "l_gripper";
palm = "l_palm";
box = "box";
table = "table";
boxSize = C.frame(box).getSize()

C.view()

tob grasps and pick-and-place over an object

#reload(manip)

C.setJointState(qHome)
C.view_raise()

C.frame(box).setRelativePosition([-.0,.3-.055,.095])
C.frame(box).setRelativeQuaternion([1.,0,0,0])

for i in range(7):
        qStart = C.getJointState()

        graspDirection = 'yz' #random.choice(['xz', 'yz'])
        placeDirection = 'z'
        place_position = [(i%3)*.3-.3, .2]
        place_orientation = [-(i%2),((i+1)%2),0.]
        info = f'placement {i}: grasp {graspDirection} place {placeDirection} place_pos {place_position} place_ori {place_orientation}'
        print('===', info)

        M = manip.ManipulationModelling(C, helpers=[gripper])
        M.setup_pick_and_place_waypoints(gripper, box, homing_scale=1e-1, joint_limits=False)
        M.grasp_top_box(1., gripper, box, graspDirection)
        M.place_box(2., box, table, palm, placeDirection)
        M.target_relative_xy_position(2., box, table, place_position)
        M.target_x_orientation(2., box, place_orientation)
        M.solve()
        if not M.feasible:
                continue

        M2 = M.sub_motion(0)
        M2.retract([.0, .2], gripper)
        M2.approach([.8, 1.], gripper)
        M2.solve()
        if not M2.ret.feasible:
            continue

        M3 = M.sub_motion(1)
#         M3.retract([.0, .2], box, distance=.05)
#         M3.approach([.8, 1.], box, distance=.05)
        M3.no_collision([], table, box)
        M3.no_collision([], box, 'obstacle')
        M3.bias(.5, qHome, 1e0)
        M3.solve()
        if not M3.ret.feasible:
            continue

        M2.play(C)
        C.attach(gripper, box)
        M3.play(C)
        C.attach(table, box)

del M

endless box pick and place with random pick and place orientations

#reload(manip)

C.delFrame('obstacle')

C.setJointState(qHome)
C.view_raise()

for l in range(20):
        qStart = C.getJointState()

        graspDirection = random.choice(['y', 'z']) #'x' not possible: box too large
        placeDirection = random.choice(['x', 'y', 'z', 'xNeg', 'yNeg', 'zNeg'])
        info = f'placement {l}: grasp {graspDirection} place {placeDirection}'
        print('===', info)

        M = manip.ManipulationModelling(C, info, helpers=[gripper])
        M.setup_pick_and_place_waypoints(gripper, box, homing_scale=1e-1)
        M.grasp_box(1., gripper, box, palm, graspDirection)
        M.place_box(2., box, table, palm, placeDirection)
        M.no_collision([], palm, table)
        M.target_relative_xy_position(2., box, table, [.2, .3])
        ways = M.solve()

        if not M.feasible:
            continue

        M2 = M.sub_motion(0)
        # M = manip.ManipulationModelling(C, info, helpers=[gripper])
        # M.setup_point_to_point_motion(qStart, ways[0])
        M2.no_collision([.3,.7], palm, box, margin=.05)
        M2.retract([.0, .2], gripper)
        M2.approach([.8, 1.], gripper)
        M2.solve()
        if not M2.feasible:
            continue

        M3 = M.sub_motion(1)
        #manip.ManipulationModelling(C, info)
        # M.setup_point_to_point_motion(ways[0], ways[1])
        M3.no_collision([], table, box)
        M3.solve()
        if not M3.ret.feasible:
            continue

        M2.play(C)
        C.attach(gripper, box)
        M3.play(C)
        C.attach(table, box)

del M

random pushes

#reload(manip)

C.frame("l_panda_finger_joint1").setJointState(np.array([.0]))

obj = box
C.frame(obj).setRelativePosition([-.0,.3-.055,.095])
C.frame(obj).setRelativeQuaternion([1.,0,0,0])

for i in range(30):
     qStart = C.getJointState()

     info = f'push {i}'
     print('===', info)

     M = manip.ManipulationModelling(C, info, ['l_gripper'])
     M.setup_pick_and_place_waypoints(gripper, obj, 1e-1, accumulated_collisions=False)
     M.straight_push([1.,2.], obj, gripper, table)
     #random target position
     M.komo.addObjective([2.], ry.FS.position, [obj], ry.OT.eq, 1e1*np.array([[1,0,0],[0,1,0]]), .4*np.random.rand(3) - .2+np.array([.0,.3,.0]))
     M.solve()
     if not M.ret.feasible:
          continue

     M1 = M.sub_motion(0, accumulated_collisions=False)
     M1.retractPush([.0, .15], gripper, .03)
     M1.approachPush([.85, 1.], gripper, .03)
     M1.no_collision([.15,.85], obj, "l_finger1", .02)
     M1.no_collision([.15,.85], obj, "l_finger2", .02)
     M1.no_collision([.15,.85], obj, 'l_palm', .02)
     M1.no_collision([], table, "l_finger1", .0)
     M1.no_collision([], table, "l_finger2", .0)
     M1.solve()
     if not M1.ret.feasible:
          continue

     M2 = M.sub_motion(1, accumulated_collisions=False)
     M2.komo.addObjective([], ry.FS.positionRel, [gripper, '_push_start'], ry.OT.eq, 1e1*np.array([[1,0,0],[0,0,1]]))
     M2.solve()
     if not M2.ret.feasible:
          continue

     M1.play(C, 1.)
     C.attach(gripper, obj)
     M2.play(C, 1.)
     C.attach(table, obj)

del M

TODOS:

• Proper execution: BotOp instead of display with C

• RRTs

• additional planar motion constraint for in-plane manipulation

• more typical manipulation constraints: camera_look_at, push_straight,

del C