In-hand manipulation

Overview

In-hand manipulation changes an object's pose while a hand, palm or finger still holds or lifts it. The arm then does not have to put the object down and regrasp it 2411.17124 §I. Four cards cover it so far, and they do not share one scope. One picks up an object too large to grasp, starting from a table 2411.16055 §I. One reorients a thin object inside the grip of a parallel gripper 2412.14899 §1. The other two move an object within a soft or deployable hand's own workspace 2411.17124 §III-A 2412.00268 §Demonstrations.

The shared difficulty is that the actuator does not control every degree of freedom of the object directly. The object moves through friction, gravity or a support it rests on. The vibration method is underactuated: one steering input cannot control both orientation and position, so orientation has to be set first 2412.14899 §4. The tilt method keeps the object from sliding away by contact with a table, a support and the palm 2411.16055 §III.

Variants

Tilt a large object against supports

A curved palm presses an object up against an inclined second support while the object rests on a first one. The palm then turns under the object and lifts it, so the object is held between the palm and the table [arxiv:2411.16055 §V]. The paper gives two tilt modes. The motion-force mode, in which one motor follows an angle and the other holds a torque, is described as gentler and as needing fewer commands 2411.16055 §V-B.

On a two-motor palm, the yoga block succeeded 5 of 5 times in both modes, and the beverage container succeeded 4 of 5 times in motion mode and 5 of 5 in motion-force mode 2411.16055 fig. 6c. On a UR3 arm with a rigid palm, a steel box failed 0 of 5 at a 90-degree support angle and succeeded 5 of 5 at 61 degrees 2411.16055 tab. I. The tests are teleoperated or programmed, open-loop and without a force/torque sensor 2411.16055 §V-A §V-D. Card: Picking by Tilting.

Suction palm with belt surfaces

DexGrip combines three soft Fin Ray fingers, each with a belt that moves the object along it, and a palm with a telescoping arm and a rotating suction cup 2411.17124 §III-A. In a palm test on 12 objects, 10 adhered and 9 succeeded overall, counting adhesion, repositioning and twisting [arxiv:2411.17124 tab. II]. The pear pivot and the cube face flip are shown, but only by eye 2411.17124 §IV-C §IV-D. Its suction model gives a maximum mass, but the accelerations and friction it needs are not stated 2411.17124 eq. 3 eq. 4. Card: DexGrip.

Tape-spring arms that roll and convey

GRIP-tape has two arms, each made of two spring-steel tape-measure strips bonded together, driven by seven motors 2412.00268 §Design p. S2. Moving the arms in the same direction rolls an object in place, and moving them in opposite directions carries it toward the base 2412.00268 §Demonstrations. A load cell in the base gives a tip-force estimate, and a 0.25 N contact threshold drives automatic gripping 2412.00268 §Sensing §Automatic. Closed-loop grip-width control keeps an ellipse in the grip, while open loop drops it 2412.00268 §Feedback fig. 8. Open-loop rotation works only for round objects 2412.00268 §Feedback. Card: Grasping and Rolling In-plane Manipulation Using Deployable Tape spring Appendages.

Vibration on a parallel gripper

An eccentric-mass motor on one finger shakes a thin object. Steering the vibration moves the object's centre of mass around the grip point, which turns the object. A radial move then sets the position 2412.14899 §2.2.4 §2.3.4. Duty-cycle vibration (50%) during translation lowered the disk's orientation error to 1.5 ± 1.02 degrees, against 3.8 ± 1.2 degrees for a constant frequency. The plate gave 2.7 against 4.5 degrees [arxiv:2412.14899 tab. 1]. Position error with the duty cycle was about 1.9 mm on both test objects 2412.14899 tab. 1. Over 10 trials per object, the credit card succeeded in 100% of tasks, the cellphone 90% and the steel ruler 70% 2412.14899 tab. 2. The pose comes from ArUco markers at 60 Hz 2412.14899 §3.2. Card: Vibration-based Full State In-Hand Manipulation of Thin Objects.

When to use which

Every number here comes from a different setup, object and trial count. Use the table to pick a family, not to rank the variants.

Implementation sketch

The vibration variant, which is the most specified. Its order is orientation first, then position, with the pose read from markers at 60 Hz.

# Orient first, then translate, with pose feedback  [arxiv:2412.14899 alg. 1]
# Input:  goal (r_g, phi_g, psi_g) in mm and degrees; pose from ArUco [§3.2]
# Output: |r - r_g| <= 1 mm and |psi - psi_g| <= 1 degree  [§3.2]

def manipulate_in_hand(goal, eps_r=1.0, eps_psi=1.0):
    # Move the centre of mass to the origin with duty-cycle vibration  [alg. 1]
    while abs(r) > eps_r:
        steer(pi)                  # theta = +-pi; sign not stated  [§2.2.4]
        duty_cycle_vibrate(50)     # Omega(t) at 50%  [§3.2]
    # Give the object a kick to start rotating about a non-zero radius  [§3.2]
    set_orbit_radius(r_c)          # r_c from the per-object limits  [tab. 2]
    kick(theta_kick, dt=0.1)       # theta_kick in [pi/6, pi/4]  [§3.2]
    # Rotate until the angle error is within tolerance  [alg. 1]
    while abs(psi - psi_g) > eps_psi:
        steer(pi)
        vibrate(168)               # constant frequency in rotation  [§3.2]
    # Return to the origin, then move radially to the goal position  [alg. 1]
    return_to_origin(eps_r)
    translate_radially(r_g, phi_g, eps_r)

The tape-spring arms do not follow this order. They roll and convey in one motion 2412.00268 §Demonstrations.

Contradictions

• Slender objects. The tilt paper says slender objects tilt more easily 2411.16055 §IV-A. The vibration paper says the steel ruler, the narrowest object it tested, succeeded least often (70%), because its narrow shape and small orbit radius make it drop during the orbit [arxiv:2412.14899 §3.3]. The mechanisms differ, and neither paper tests the other's claim.
• Friction. The tilt paper attributes the palm's stability to high palm friction 2411.16055 §V-D. It plans with friction values of 0.1 and 0.2 at that contact, and the hardware's real values are not stated 2411.16055 fig. 3. The tape-spring paper names belt-to-finger friction as a limit 2412.00268 §V and does not measure it.
• Rotation range. The suction-palm paper's abstract claims full 360-degree rotation about all three axes. Its palm rotates about one axis, and the paper reports no angle range per axis 2411.17124 §IV-B.
• Measures. No two papers report the same quantity. Orientation error is in degrees for the vibration paper 2412.14899 tab. 1 and tip error in mm for the tape-spring paper 2412.00268 §Kinematics. Success counts use different objects and trial counts in each card.

Papers

• Picking by Tilting: In-Hand Manipulation for Object Picking using Effector with Curved Form: tilts a large object against two supports, then tucks a curved palm under it.
• DexGrip: Multi-modal Soft Gripper with Dexterous Grasping and In-hand Manipulation Capacity: soft gripper with belt-driven fingers and a telescoping suction palm.
• Grasping and Rolling In-plane Manipulation Using Deployable Tape spring Appendages: two tape-spring arms that roll and convey objects in a plane.
• Vibration-based Full State In-Hand Manipulation of Thin Objects: an eccentric-mass motor on one finger sets position and orientation of thin objects.