Skip to content

4. A commanded hand pose does not establish a successful pick

Before: frames and trajectory execution. Build: transform a grasp into the world, estimate one simple contact-force requirement, and distinguish hand motion from object retention.

Name which frame each grasp belongs to

Suppose perception estimates an object's world pose T_WO. A grasp proposal specifies a desired gripper pose relative to that object, T_OG. The desired world gripper pose is:

T_WG = T_WO × T_OG

The notation means that T_AB converts coordinates expressed in B into A. It also describes B's pose relative to A. If a dataset instead stores the object's pose relative to the gripper, that is T_GO; invert it before this composition. Names such as grasp_pose do not establish the convention.

Our fixture supports translation in x/y/z and rotation about z only; it does not implement arbitrary roll and pitch. Let the world object pose be:

translation = (0.6, 0.2, 0.0) m
yaw = +90 degrees
object-relative gripper translation = (0.1, 0.0, 0.05) m
relative yaw = 0

Rotate the relative x offset into world y, then add the world translation. The desired gripper pose is (0.6, 0.3, 0.05) m at +90°. Directly adding the translations would incorrectly produce (0.7,0.2,0.05).

Guided variation: change only the object yaw to zero. The world gripper translation then becomes (0.7,0.2,0.05). The object-frame grasp specification has not changed; its world coordinates have.

A grasp proposal still needs a feasible approach

A useful execution sequence separates object localization, pregrasp approach, contact acquisition, closure, lift, transport, placement, and release. Each stage has a distinct completion condition. Reaching the pregrasp pose does not mean the fingers have made the intended contact.

Before execution, check inverse kinematics, joint limits, the gripper's opening, approach clearance, and the geometry of the hand and object. Once the object is held, its geometry moves with the robot and must be included during transport. A target reachable by the tool center can still cause a wrist or carried-object collision.

These checks connect a vision or VLA output to the motion and actuator contracts from the previous lessons. A learned model's confidence is not a substitute for an explicit geometric or execution condition.

One contact calculation with visible assumptions

For an ideal symmetric two-finger grasp, let each finger apply normal force N against opposing vertical faces. Assume Coulomb friction coefficient mu, a stationary upright object, equal load sharing, and no relevant tipping moment. Each contact can supply tangential force up to mu × N, so vertical holding requires:

2 × mu × N >= m × g
N_required = load_factor × m × g / (2 × mu)

For mass 0.5 kg, mu=0.4, and illustrative load factor 1.5, the estimate is 9.196875 N per finger. It is not the total of both normal forces. Halving mu doubles the required normal force. Faster upward acceleration increases the vertical load; a tipping moment requires a wrench analysis beyond this scalar calculation.

The Modern Robotics friction lesson develops contact/friction-cone models. Our assumed coefficient and load factor are teaching values, not measured material properties or an actuator rating. Increasing squeeze force can exceed a gripper's continuous capability or damage an object; those are additional limits, not effects represented by this formula.

Does this establish force closure? No. Resisting this particular vertical load is weaker than resisting arbitrary disturbance wrenches. Two ideal point contacts in 3-D do not generally resist a torque about their connecting line. Contact locations, normals, friction, and contact model matter. See the primary force-closure lesson.

Measure the object relative to the hand

If an object remains rigidly held, its pose relative to the gripper stays approximately constant:

T_GO_before = inverse(T_WG_before) × T_WO_before
T_GO_after  = inverse(T_WG_after)  × T_WO_after

Compare relative translation and orientation using appropriate tolerances. The supplied yaw-only fixture uses Euclidean translation error and the shortest wrapped yaw difference. A full 3-D implementation needs a rotation metric on SO(3); subtracting quaternion components is not a general angular error.

The paired cases use the same commanded and final hand pose:

Observation after the hand rises 0.2 m Object lift Relative translation change Result
Object rises with the hand 0.2 m approximately 0 m Retained and lifted
Object stays on the table 0 m 0.2 m Failed pick

The demonstration defines lift success as at least 0.15 m object lift, relative translation change at most 0.01 m, and yaw change at most 0.05 rad. Keeping everything still has zero retention error but fails the lift requirement.

These pose comparisons rely on valid object observations, frame calibration, and timestamps. If the object becomes occluded, label retention unknown or use another validated observation; do not reuse the last visible pose as proof of continued contact. The CPU fixture has exact synthetic poses and no contact dynamics.

Extend success to the actual task

A pick-and-place task needs additional criteria: the object reaches the allowed placement region, is released, and remains there for a declared interval with acceptable motion. The hand leaving the goal is not proof that the object was released. A camera image alone may be ambiguous about whether the object is supported, wedged, or still attached.

Use stage-specific outcomes such as approach_failed, contact_not_confirmed, retention_lost, placement_outside_region, and release_unconfirmed. Keep an overall task-success metric too. This helps explain why a model with excellent action prediction or numerical parity can perform worse in closed-loop tasks.

Independent transfer: add three timestamped post-release object observations. Require all three to lie inside a chosen placement region with speed below a declared limit. Include a case that briefly enters the region and then falls out. Use actual elapsed time, not just sample count, if observation rates differ.

Run python3 labs/control_lab.py grasp from this course directory. The implementation supplies transform composition, relative-pose error, the two-contact force estimate, and the lift predicate. Complete the lesson with the hand calculation, paired outcomes, and a clear account of which contact, perception, and release assumptions remain untested.