⚡ electrical engineering

Grasp Execution Without a Planner: Configuration-Space Grasp Distance Fields with Certified Safety & Guaranteed Quality

This paper introduces Grasp Distance Fields (GDFs), a planner-free control framework that executes robust and safe grasps by following a smooth configuration-space distance field filtered through a safety-critical quadratic program, achieving high success rates and quality retention in dynamic environments without the need for trajectory replanning.

Original authors: Clinton Enwerem, John S. Baras, Calin Belta

Published 2026-08-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Clinton Enwerem, John S. Baras, Calin Belta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where robots are like clumsy toddlers trying to pick up a toy. They can see the toy, but the moment they reach out, their hand might bump into a chair, or the toy might shift slightly, causing the robot to drop everything and start over. This is the daily struggle of "multifingered grasping" in robotics. For decades, the standard solution has been a two-step process: first, a computer brain plans a perfect, collision-free path to the object (like drawing a map), and then the robot tries to follow that map. But if the world changes even a tiny bit—a person walks by, or the object slides—the map becomes useless, and the robot has to stop, panic, and draw a new map. This "plan-then-track" method is slow and fragile. Scientists want robots that can react instantly, like a human grabbing a coffee cup while walking through a crowded room, without needing to stop and think about every single step.

This paper introduces a new way for robots to grab things called "Grasp Distance Fields" (GDFs). Instead of drawing a specific map or picking one "best" way to grab an object, the robot is given a smooth, invisible landscape of "goodness." Imagine standing in a foggy field where the ground slopes gently downward toward every possible way you could successfully grab a cup. The robot's job is simply to roll downhill. It doesn't need to decide which cup to grab or how to grab it beforehand; it just follows the slope of the field. If an obstacle appears, the robot naturally steers around it because the "slope" changes to guide it safely. The authors prove mathematically that this method keeps the robot safe from crashing into itself or the environment, and they show that even if the robot's hand is a bit wobbly, it can still hold on tight enough to lift the object. They tested this on a robot arm and a humanoid robot (the Unitree G1), showing it could successfully grab 46 out of 50 different objects, even in messy, crowded scenes, without ever needing to stop and replan.

The "Rolling Downhill" Robot

Most robots that try to pick things up work like a GPS navigation system that gets confused if you take a wrong turn. They calculate a perfect route to a specific spot, and if you move the destination even an inch, the robot has to stop, recalculate the whole route, and start over. This paper proposes a different approach: instead of a GPS, give the robot a "feeling" of where the goal is, like a ball rolling down a hill.

The researchers created a mathematical "field" (a kind of invisible map) over all the possible positions a robot's arm and hand could be in. In this field, every point represents a different pose. The "height" of the field at any point tells the robot how far it is from a successful grab. The robot's controller is programmed to simply follow the steepest downhill path (the negative gradient) toward the lowest point. Because this field is smooth and covers all possible good grasps at once, the robot doesn't need to pick a single "winner" beforehand. It just rolls toward the nearest valley. If the robot is blocked by a wall, the field naturally guides it around the obstacle, just like water flowing around a rock.

No More "Stop and Think"

The biggest problem with traditional robots is the "replanning" step. If the object moves, the robot freezes. This paper's method eliminates that freeze entirely. The robot reacts in real-time, millisecond by millisecond. The authors call this "reactive execution."

To make sure the robot doesn't crash while rolling downhill, they added a "safety filter." Think of this as a very strict guardian angel. The robot might want to roll straight toward the object, but if that path would cause it to hit a wall or its own elbow, the guardian angel steps in. It uses a mathematical tool called a "Control Barrier Function" (CBF) to gently nudge the robot's path just enough to stay safe, without stopping the progress. The robot solves a tiny math puzzle (a quadratic program) in just 0.09 milliseconds—faster than a human eye blink—to decide the safest move.

The "Hysteresis" Trick: Knowing When to Let Go

One tricky part of grabbing things is knowing when to stop moving and actually hold on. If a robot tries to grab a cup, its fingers might touch the cup, but if it keeps pushing, it might slip or drop it. The paper introduces a "mode switch" with a feature called "hysteresis."

Imagine a door with a heavy spring. You have to push it open past a certain point to get it to stay open, but you don't have to push it all the way back to close it. The robot uses this same logic. It has a "Reach" mode, a "Close" mode, and a "Hold" mode. To switch from reaching to closing, the robot must get very close to the object. But to switch from holding back to closing (if it slips), it has to move farther away than it was when it started holding. This prevents the robot from getting confused and switching modes back and forth rapidly (chattering) when it's right on the edge of touching the object.

Proving It Works (Even When Things Go Wrong)

The authors didn't just guess this would work; they proved it mathematically and tested it extensively.

The Math: They proved that if the robot starts in a safe area, it will always stay in a safe area, no matter what obstacles appear. They also proved that the "quality" of the grab (how tightly it holds) won't drop below a safe level once it starts holding.
The Simulation: They tested their system on a 7-joint robot arm with an 11-joint hand, and also on a full humanoid robot (the Unitree G1).
The Results: In a cluttered room with 50 different objects (from simple shapes to complex household items and tricky "adversarial" shapes designed to fool robots), the robot successfully grabbed and lifted 46 of them.
The "Quality" Check: They found that the grasps the robot actually performed retained a median of 94% of the theoretical "quality margin" they were designed for. This means the robot didn't just grab the object; it grabbed it securely.

What It Doesn't Do (Yet)

It's important to know what this paper doesn't claim.

It's not magic: The robot still needs to know where the object is. If the object moves too fast or the robot's sensors are wrong, the system can get stuck. The authors found that in some very tight spots, the robot could get "trapped" in a local valley where it can't move forward or backward without hitting something. They call this a "trapped equilibrium."
It's not a hardware demo: All the results shown here are from computer simulations. The robot never actually lifted a real cup in the real world in this study. The authors note that while the math looks solid, testing on real hardware with real friction and vibrations is the next step.
It doesn't solve everything: If an obstacle blocks every possible path to the object, the robot will stop safely short of the object rather than crashing. It won't magically find a way through a wall.

Why This Matters

This paper offers a new way to think about robot control. Instead of trying to predict the future and draw a perfect map, it gives the robot a simple, robust rule: "Roll downhill, but don't hit anything." This makes the robot much more resilient to mistakes and changes in the environment. By proving that this method is mathematically safe and works in complex simulations, the authors provide a strong foundation for building robots that can work safely alongside humans in messy, unpredictable real-world environments. The fact that the same code worked on a simple arm and a complex humanoid robot suggests this approach could be a universal key to making dexterous robots a reality.

Technical Summary: Grasp Execution Without a Planner

Problem Statement
Standard multifingered grasp execution relies on a "plan-then-track" architecture: a generator proposes candidates, a selector chooses one, and a motion planner computes a collision-free trajectory to that specific pose. This approach is fragile; execution-time uncertainties in object pose or system perturbations can invalidate the planned trajectory, necessitating a costly and latency-inducing replanning step. Furthermore, existing reactive methods often lack formal safety guarantees or fail to integrate grasp quality metrics (such as force closure) directly into the execution loop. The paper addresses the need for a reactive, planner-free execution framework that guarantees safety, handles dynamic obstacles, and maintains grasp quality without discrete selection steps.

Methodology
The authors propose Grasp Distance Fields (GDFs), a control framework that replaces discrete trajectory planning and grasp selection with a continuous, smooth feedback law operating directly in the configuration space of the coupled arm-hand system.

Grasp Distance Fields (GDFs):
- Instead of selecting a single grasp, the system defines a target set $\mathcal{G}$ of pre-grasp configurations derived from a candidate set.
- A softmin distance field $d_G(q)$ is constructed over this set in the configuration space. This smooth approximation of the minimum distance to any candidate eliminates the non-smoothness (and resulting chatter) associated with Voronoi boundaries between candidates.
- The nominal control command follows the negative gradient of this field ( $v_{nom} = -k \nabla d_G$ ). The softmax weights of the softmin function implicitly perform continuous candidate selection, always directing the system toward the nearest candidate in joint space without a discrete decision step.
Certified Safe Execution (CBF-CLF QP):
- The nominal command is filtered through a Quadratic Program (QP) at every control step.
- Safety (CBF): The QP enforces Control Barrier Function (CBF) constraints for self-collision, workspace limits, object clearance, and dynamic obstacle avoidance. This ensures the safe set is forward invariant.
- Convergence (CLF): Control Lyapunov Function (CLF) constraints drive the system toward the target set. To handle conflicts between safety and convergence, the QP introduces explicit slack variables to record blocked progress rather than violating safety.
- Underactuation: The QP includes equality constraints to enforce the kinematic coupling of underactuated hands.
Hybrid Mode Switching:
- Since smooth fields cannot inherently model the discontinuity of contact admission, the system employs a hybrid automaton with four modes: Reach (R), Close (C), Hold (H), and Lift (L).
- Transitions are governed by hysteresis-based guards on contact indicators and duration timers to prevent chatter.
- In the Hold mode, the system enforces a wrench-quality barrier. This constraint ensures that the risk-adjusted force-closure margin ( $\varepsilon^{(\beta)}$ ) of the realized grasp does not decay beyond a prescribed tolerance from its value at the onset of the hold.
Wrench-Certificate Bridge:
- The framework connects joint-space distance fields to wrench-space quality certificates. It admits only grasps where the pre-grasp descriptor satisfies a risk-adjusted margin threshold.
- During execution, it monitors the realized margin against the stored descriptor's margin, guaranteeing that the executed grasp remains within a tolerance of its synthesized quality.

Key Contributions
The paper makes five specific contributions:

Grasp Distance Field: Definition of a smooth softmin distance field over a discrete grasp candidate set in configuration space, establishing bounds on its approximation error ( $\log N / \rho$ ) and proving that its gradient is a convex combination of candidate gradients, thereby eliminating discrete selection.
Certified Safe Execution Law: A stationary feedback law filtered through a combined CBF-CLF QP that enforces safety (self-collision, obstacles, workspace) and records convergence failure via slack variables. It proves forward invariance of the safe set.
Hysteresis-Based Contact Switch: A hybrid system structure that manages the transition from approach to contact and hold, using hysteresis to ensure robust mode transitions without replanning.
Wrench-Certificate Bridge: A constraint that maintains the risk-adjusted force-closure margin within a tolerance of the hold-start value, connecting the joint-space controller to wrench-space quality guarantees.
Full-Description Verification: Numerical verification on a 7-DOF arm with an 11-joint underactuated hand and a Unitree G1 humanoid, demonstrating the controller's ability to navigate clutter, handle dynamic obstacles, and execute grasps without modification across different platforms.

Experimental Results
The controller was evaluated in kinematic simulation on a fixed-base arm-hand system and a Unitree G1 humanoid, both using the same underactuated hand.

Success Rate: The controller successfully reached, grasped, and lifted 46 out of 50 objects (spanning primitives, household items, and adversarial meshes) in cluttered and dynamic scenes.
Quality Retention: Across successful lifts, the executed grasps retained a median of 94% of their synthesized quality margin.
Safety: The controller avoided all obstacles and self-collisions. In cases where the approach was blocked, the system stopped safely short of the object without collision.
Quality Filtering: When a wrench-quality constraint was enforced, the system correctly rejected a grasp (the "cube" trial) that a quality-neutral controller would have attempted to lift, preventing a failure where the object would have fallen under gravity.
Performance: The QP solved in 0.09 ms per step (with a 20 ms control period), and the full step including collision distances took 2 ms.
Dynamic Obstacles: The controller successfully navigated around moving obstacles without replanning, reacting to pose changes by re-evaluating the same field map.

Significance and Claims
The authors claim that this work provides the first planner-free, reactive grasp execution framework that offers certified safety and guaranteed quality retention without a separate grasp selection step.

Reactivity: By removing the replanning step, the system reacts instantly to object movement or state disturbances.
Safety: The use of Control Barrier Functions provides formal guarantees that the system remains within safe sets (no collisions) even when convergence is blocked.
Quality: The integration of risk-adjusted force-closure margins directly into the execution loop ensures that the robot does not lift objects it cannot securely hold, bridging the gap between grasp synthesis and execution.
Generality: The controller operates on a single stationary feedback law that requires no modification to run on different robotic platforms (e.g., fixed-base arm vs. humanoid), provided the robot description (kinematics and underactuation) is correctly specified.

The paper acknowledges limitations, including the inheritance of local minima from potential field methods (trapped equilibria in dense clutter) and the reliance on kinematic simulation for verification, noting that dynamic wrench certificates remain an open question for hardware implementation. However, it asserts that the proposed method significantly advances the state of the art by unifying safety, quality, and reactivity in a single, planner-free control law.