Grasp Execution Without a Planner
The standard architecture for multifingered grasp execution plans a collision-free trajectory to a selected grasp pose and tracks it with a feedback law, and a moved object or a disturbed state forces a costly replanning step. We remove the replanning step entirely and represent the target as a set of grasp configurations, over which we define a smooth softmin distance field in the configuration space of the arm-hand system. Our controller follows the field's negative gradient with a stationary feedback law while a CBF-CLF quadratic program keeps every step inside the safe set.
We prove that the softmin distance tracks the true set distance within an explicit gap, that its gradient is a convex combination of unit-norm candidate directions with a resulting speed bound, and that the filtered closed loop renders the safe set forward invariant. Since no smooth field captures the switching that hand-object contact introduces, a mode switch with hysteresis replaces object collision avoidance with contact admission at a realized pregrasp configuration. A wrench-quality barrier constraint then keeps the realized grasp's force-closure margin within a prescribed tolerance of its value at the start of the hold, and we evaluate every executed grasp under the same risk-adjusted margin that certified its descriptor.
We evaluate in kinematic simulation on two platforms, a fixed-base arm and a Unitree G1 humanoid, both fitted with the same underactuated multifingered hand. Our controller safely reaches for, grasps, and stably lifts 46 of 50 objects spanning primitive, household, and adversarial classes, the executed grasps keep a median 94 percent of their synthesized quality margin, and the quadratic program solves in 0.09 ms per 20 ms control step. The recorded per-step softmax weights confirm that our controller executes whichever candidate the field places nearest, with no separate selection step.
The target is a set of certified pregrasp configurations rather than a single pose. Over pregrasp configurations $q^{\text{pre}}_i$, the field is a log-sum-exp softmin under a diagonal metric that weights the arm above the hand, $d_G(q) = -\tfrac{1}{\rho}\log \sum_i \exp(-\rho\,\lVert q^{\text{pre}}_i - q\rVert_\Lambda)$. It tracks the true set distance within $\log N/\rho$, its gradient is a convex combination of unit directions, and the negative gradient points toward whichever candidate lies nearest with no selection step. A control-barrier quadratic program with a linear class-$\mathcal{K}$ rate filters the nominal command. An admission test eliminates any candidate whose pregrasp or closure violates a barrier constraint, and the target set therefore contains only certified grasps.
A softmin over certified pregrasp configurations in joint space, smooth, with a bounded gradient and an explicit gap to the hard minimum.
A CBF quadratic program over self-collision, workspace, object, and obstacle constraints, with the dependent-joint constraints of the underactuated hand as equalities.
A mode structure with hysteresis that admits contact finger by finger, the single event a smooth field cannot express.
We evaluate the executed grasp under the same risk-adjusted force-closure margin that certified its descriptor, and we report the share of the synthesized margin the executed grasp keeps.
A Unitree G1 humanoid runs the identical controller, and only the robot description and scene data change.
The panel below plays back a recorded trial. The tabs switch the trial, the slider scrubs the timeline, and the readout reports the field value, the minimum barrier margin, and the number of contacts across the four execution modes, reach, close, hold, and lift.
Top-down map of the workspace. The palm follows the grasp field, the shaded well around the target, and the bar at the right edge shows the lift height.
We execute certified grasp records over 50 objects, four primitive shapes, 17 YCB household objects, and 29 adversarial EGAD meshes, each behind a rectangular obstacle placed next to its approach path. Our controller completes the reach-grasp-lift sequence on 46, 37 of the 48 evaluated grasps satisfy the risk-adjusted margin $\varepsilon^{(\beta)} \ge 0$ at $\beta = 0.9$, and 39 satisfy single-friction force closure.
| Object class | Objects | Lift | $\varepsilon^{(\beta)} \ge 0$ | Force closed | Margin ≥ 0.3 |
|---|---|---|---|---|---|
| Primitives | 4 | 4 | 3 | 3 | 2 |
| YCB household | 17 | 14 | 12 | 12 | 10 |
| EGAD adversarial | 29 | 28 | 22 | 24 | 18 |
| All | 50 | 46 | 37 | 39 | 30 |
Every failure is a failure of convergence, never of safety. In two trials, the hand never attains three contacts within the horizon and the trial ends in close, in the other two, our controller reaches hold but never satisfies the lift guard, and the quadratic program remains feasible at every step of every trial.
The reference scene admits no unobstructed approach. The first playback tab and video show a 6 by 6 by 50 cm obstacle pair with the object behind it, and the gap between the obstacles is the only route. The palm deviates 20.5 cm around the obstacle, both obstacle constraints stay positive from the start pose onward with a minimum of 6.7 mm, and the executed grasp keeps 91 percent of its descriptor's force-closure margin. We map the boundary of the trapped region with a scene-placement search, a 50 cm column on the approach path leaves the closed loop trapped in every trial, and in the reported scene, the trajectory deviates around a first obstacle before reaching the obstacle on its line.
Our controller handles a moving obstacle with no new constraints. It updates the obstacle poses at every control step as the 5 by 5 by 45 cm obstacle crosses the approach path in the second tab. The palm deviates up to 19.1 cm, 14.9 cm by the point-to-path measure, and surface clearance stays positive with a minimum of 8.4 mm. We also use this case to measure the effect of the time-derivative term the program does not include. With no obstacle-velocity contribution, an active constraint falls below zero by at most the approach speed over the barrier rate, 6.6 mm measured against the 10 mm bound and the 1.5 cm margin.
The same controller runs on a humanoid. The third tab and video play back a Unitree G1 with the same hand merged at the right wrist, three torso joints joining the seven-joint arm chain, and a grasp record synthesized for that platform. Our controller architecture is identical, and only the description and scene data change. A 4 by 4 by 36 cm obstacle stands on the palm-to-object line, the palm deviates up to 9.7 cm, and the object rises 12 cm with the torso rotation under 18 degrees.
arXiv link coming soon.
@online{enweremGraspDistanceFields2026,
title = {Grasp Execution Without a Planner: Configuration-Space Grasp Distance Fields with Certified Safety \& Guaranteed Quality},
author = {Enwerem, Clinton and Baras, John S. and Belta, Calin},
year = {2026},
eprint = {},
eprinttype= {arxiv},
eprintclass = {cs.RO},
doi = {},
url = {},
pubstate = {prepublished},
keywords = {Computer Science - Robotics, Computer Science - Systems and Control},
}