Grasp Planning under Uncertainty

A candidate approach specifies a travel direction, palm roll, finger preshape, and set of joints to close. For a given hand and object, the contact planner enumerates these commands, simulates the approach and finger closure, and ranks the admitted configurations by modeled wrench capacity. The output has the contact geometry and scoring samples needed to reproduce the ranking. The calling application evaluates arm reachability and plans and executes the arm trajectory.

Set the Hand Frame

The hand configuration defines a palm-frame approach axis, a second axis along the fingers, and a grasp center. The approach axis points toward the object. The planner projects the finger axis onto its orthogonal plane to define roll, so the two input axes must be nonzero and nonparallel. The grasp center \(g\) is a hand-fixed point near the region where the fingers close. If \(R\) denotes the planned palm-to-object rotation, \(c\) the object center of mass, \(d\) the unit travel direction, and \(s\) the standoff distance, the starting palm translation is \(c-sd-Rg\). Thus the grasp center begins \(s\) millimeters from \(c\) along the approach line. The RealHand L6 walkthrough shows the axes and coordinates for an example model. Explicit preshapes can also be sampled from a low-dimensional posture subspace before running the planner. See Ciocarlie and Allen for dexterous grasp synthesis over such subspaces.

Specify Candidate Commands and Contact Constraints

The planner advances the open hand in discrete increments along \(d\) until an increment would cause collision. It starts closure at the last collision-free placement, with the requested clearance subtracted from the distance traveled. GraspIt! advances the selected finger coordinates under autograsp. The planner records closures with enough object contacts and applies the configured contact-region predicate. Support and keep-out predicates inspect the final nominal hand vertices. This provides a geometric filter for the candidate bank. The calling application verifies continuous hand and arm clearance during execution.

Setting

Meaning

approaches

Explicit rows [dx, dy, dz, roll_deg] in the object frame. In their absence, n_directions gives the number of Fibonacci-sphere directions and n_rolls the number of equally spaced rolls per direction.

preshapes

Starting joint vectors in exported DOF order, in radians. Every direction and roll is evaluated with every specified preshape and closing set.

closing_sets

Lists of zero-based DOF indices. Coordinates outside the chosen list remain at preshape during closure.

standoff_mm, approach_step_mm, clearance_mm

Initial approach distance, travel increment, and retreat from the last collision-free placement, all in millimeters.

min_contacts

Minimum number of object contacts required at closure.

support_normal, support_offset_mm

Support half-space in the object frame. The planner normalizes the vector and admits a nominal closure if every hand vertex lies at least the specified offset along it.

contact_region_mm

Allowed contact box in object coordinates, given by its lower and upper corners as [[xmin, ymin, zmin], [xmax, ymax, zmax]].

keepout_boxes_mm

Excluded boxes in object coordinates, each given by its center and half extents. The filter rejects a nominal closure if any hand vertex lies strictly inside an excluded box.

Select a Wrench Score

The default score uses Ferrari–Canny epsilon to quantify the worst direction of resistance to a normalized wrench. For task-specific synthesis, supply task_forces_mm as rows [px, py, pz, fx, fy, fz]. The first three entries give the point of application in the object frame in millimeters. The final entries give the contact force direction required to balance the external load. The planner normalizes each nonzero force vector before constructing its wrench. For instance, [0, 0, 0, 0, 0, 1] represents upward support through the object-frame origin. The score is the greatest magnitude supported across the supplied directions.

Model Approach and Friction Uncertainty

The planner closes the hand once at the nominal approach and again for each of n_pose_samples perturbations. The parameters pose_sigma_mm and pose_sigma_deg set the componentwise standard deviations of isotropic Gaussian translation and rotation-vector errors. The translation, rotation, and travel direction are expressed in the object frame. The seed and command together determine the sampling stream. Each friction coefficient in friction_samples is applied to each realized contact set when constructing its wrench hull. The friction grid has equal weights, and a failed perturbed approach or closure contributes zero across its friction row.

The planner ranks candidates by the lower-tail average over the lowest-scoring tail_mass fraction of pose–friction combinations. This score favors capacity in weak scenarios even when competing approaches have similar means. Set tail_mass=1 for mean ranking. Set n_pose_samples=0 and provide one friction coefficient for ranking at the specified nominal coefficient. The theory guide defines both quality measures and the empirical tail calculation.

Interpret Candidate Records

The CLI orders the JSON records by decreasing score and writes the nominal palm placement T_object_hand_mm, closed coordinates dofs_rad, and per-contact positions and inward normals. The homogeneous transform maps palm coordinates into the object frame. Position arrays and transform translations use millimeters, and DOFs use radians. Preshape, travel direction, roll, and closing indices identify the approach.

epsilon_samples and task_samples contain pose rows and friction columns. The first row is nominal. nominal_epsilon and nominal_task read the middle column of the nominal row, using the upper middle column for an even grid. expected_epsilon and expected_task are means over the corresponding arrays. epsilon_cvar stores the epsilon lower-tail mean even in task mode, where score instead stores the task lower-tail mean. native_epsilon comes from the simulator at the nominal closure and uses the simulator’s wrench normalization. The top-level rejections field counts causes of nominal rejection. See MuJoCo model export for converting the palm and joint targets to the source model.

Planned Grasps and Simulated Execution

Each \(\varepsilon\) value is the Ferrari–Canny score returned by GraspIt! for the planned closure. The L6 panel shows the planned closure on the insertion key. The other panels show final states of recorded lift-and-hold simulations in MuJoCo. In those panels the score belongs to the GraspIt! plan used to initialize each execution.