💻 computer science

2D and 3D Grasp Planners for the GET Asymmetrical Gripper

This paper introduces GET-2D-1.0 and GET-3D-1.0, two grasp planners for the GET asymmetrical gripper that utilize 2D RGB-D and 3D mesh-based approaches respectively, demonstrating significant performance improvements over baselines while highlighting a trade-off between the high speed of the 2D method and the marginal accuracy gains of the more computationally expensive 3D method.

Original authors: Andrew Goldberg, Ethan Ransing, Anton Kourakin, Cael Magner, Edward H. Adelson, Ken Goldberg

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew Goldberg, Ethan Ransing, Anton Kourakin, Cael Magner, Edward H. Adelson, Ken Goldberg

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 robot trying to pick up a weirdly shaped object, like a toy car with a handle or a rock with a hole in it. For decades, robot hands have been like standard kitchen tongs: two straight, parallel jaws that pinch from opposite sides. This works fine for simple boxes, but it's like trying to pick up a slippery bar of soap with just two fingers; if you try to lift it off-center, it twists and slips right out of your grip.

This paper introduces a new kind of robot hand called the GET gripper and two "brains" (planners) that teach it how to hold things securely.

The New Hand: The "Asymmetrical Gripper"

Think of the old robot hands as having two identical fingers. The new GET gripper is different. It has a V-shape on one side (like a wide, open palm) and a single narrow finger on the other.

The Analogy: Imagine trying to hold a wrench. If you use two flat fingers, it might spin. But if you use a V-shaped cradle on one side and a single finger on the other, the object gets "locked" in place. The V-shape acts like a stabilizing lever, preventing the object from twisting even if you lift it from the side. This design allows the robot to grab things with three points of contact instead of just two, making the grip much stronger and more stable.

The Two "Brains" (Planners)

The paper presents two different ways to tell this new hand where to grab.

1. GET-2D-1.0: The "Fast Sketch Artist"

This planner is like a quick sketch artist looking at a flat photo of an object.

How it works: It takes a single picture from above (like a bird's-eye view), traces the outline of the object, and turns it into a simple polygon (a shape made of straight lines).
The Strategy: It doesn't just guess. It uses a special math trick (called the Ferrari-Canny metric) to simulate thousands of "what-if" scenarios. It asks, "If I grab here, will the object slip if I shake it?" It specifically looks for opportunities to stick the narrow finger into holes or rest the wide V-shape against edges.
The Result: It is incredibly fast (taking less than a second) and is surprisingly good at finding strong grips.

2. GET-3D-1.0: The "Slow 3D Architect"

This planner is like a detailed architect building a 3D model of the object.

How it works: Instead of just a flat outline, it uses a full 3D mesh (a digital 3D model) of the object. It shoots "lasers" (rays) from the gripper's fingers to see exactly where they will touch the object in 3D space.
The Strategy: It does a much more precise calculation of how the fingers will press against the object's curves and corners.
The Result: It finds slightly better grips than the 2D version, but it is very slow, taking about 17 seconds to plan a single grab.

The Big Test: The "Shake and Pull"

To see which method worked best, the researchers put the robot to the test with 10 different tricky objects (like a rock climbing hold, a tape roll, and a gearbox). They compared their new planners against a basic method that just tries to grab the object's "bounding box" (an imaginary rectangle around the object).

They used two tests:

The Lift: Can it pick the object up without dropping it?
The Shake: The robot shakes the object back and forth violently to see if it slips.
The Pull: A force gauge measures how hard they have to pull to twist the object out of the gripper.

What They Found

The 2D Planner (The Fast One): It was a huge winner. Compared to the basic "bounding box" method, it improved success rates by over 40%. It was much better at lifting objects, surviving the shake test, and resisting being pulled out. It proved that you don't need a slow, complex 3D model to get a great grip; a smart 2D strategy works wonders.
The 3D Planner (The Slow One): It was the most successful, lifting 100% of the objects and surviving the shake test 95% of the time. However, because it takes 17 seconds to think, it's not practical for a robot that needs to move quickly.
The Trade-off: The 2D planner is fast and nearly as good as the 3D one. The 3D planner is slightly more robust but too slow for real-time use.

The Bottom Line

The paper shows that by combining a new, cleverly shaped hand (the asymmetrical GET gripper) with a smart, fast planning algorithm (GET-2D-1.0), robots can pick up difficult, oddly shaped objects much more reliably than before. It's like upgrading from a clumsy pair of tongs to a specialized tool that knows exactly how to cradle a fragile or slippery item, all while thinking fast enough to keep up with a busy workday.

1. Problem Definition

Traditional industrial parallel-jaw grippers typically utilize symmetric narrow jaws. While mechanically simple, this design relies on two antipodal contact points, creating a pivot axis with limited resistance to external torques. This leads to instability when manipulating off-center payloads or performing tasks requiring torque resistance (e.g., sweeping, drawing). Conversely, symmetric wide jaws increase torque resistance but often cause unintentional collisions and occlude the field of view for overhead vision systems.

The GET (Grasp Everything) gripper, introduced by Burgess and Adelson (2025), addresses these limitations with a 1-DOF asymmetric design featuring a "two-against-one" configuration: a wide V-shaped jaw opposing a single narrow jaw. This geometry enables three points of contact, forming a stabilizing lever arm against torque disturbances. However, existing grasp planners (designed for symmetric grippers, suction cups, or multi-finger hands) do not exploit this unique geometry. The paper aims to develop and evaluate algorithmic planners specifically for the GET gripper to maximize grasp robustness.

2. Methodology

The authors propose two distinct grasp planning algorithms: GET-2D-1.0 (fast, image-based) and GET-3D-1.0 (precise, mesh-based). Both utilize Force Closure and the Ferrari-Canny metric (ε-distance) to evaluate grasp quality.

A. GET-2D-1.0 (2D Planner)

Input: Single-view RGB-D image from a wrist-mounted camera.
Pipeline:
1. Segmentation: Converts the image to a binary mask via color thresholding.
2. Polygon Approximation: Extracts contours and simplifies them into a polygon using the Teh-Chin and Ramer-Douglas-Peucker algorithms.
3. Sampling Strategy: The gripper is modeled as three circles (two for the wide jaw, one for the narrow). The planner uses three complementary sampling strategies:
  - Edge Sampler: Places the wide jaw normal to polygon edges.
  - Vertex Sampler: Straddles polygon vertices with the wide jaw.
  - Hole Sampler: Inserts the narrow jaw into holes within the object polygon.
4. Evaluation: Candidates are evaluated by sampling 2D translational perturbations (Gaussian noise). Metrics (Force Closure Rate, contact count, and average Ferrari-Canny metric) are averaged across perturbations to ensure robustness.
Output: The highest-ranked grasp $(p, \phi, z)$ .

B. GET-3D-1.0 (3D Planner)

Input: RGB-D image + Pre-acquired 3D mesh of the object.
Pipeline:
1. Alignment: Aligns the mesh to the point cloud using Iterative Closest Point (ICP).
2. Sampling: Randomly samples mesh faces (weighted by area) to orient the wide jaw.
3. Ray-Tracing: Casts rays from the gripper contact surfaces to determine intersection points with the mesh.
4. Optimization: An optimizer adjusts the gripper's rotation to minimize the imbalance in contact distances between the two wide jaw fingers, ensuring a stable three-point contact.
5. Evaluation: Similar to the 2D planner but operates in 3D space, computing force closure and Ferrari-Canny metrics on the mesh.
Output: The optimal 6-DoF grasp pose.

3. Key Contributions

GET-2D-1.0: A fast, single-view 2D grasp planner that leverages the GET gripper's unique geometry. It introduces a novel "Hole Sampler" to utilize object concavities, which standard bounding box methods miss.
GET-3D-1.0: A mesh-based 3D planner that efficiently samples candidates and uses ray-tracing and gradient descent to optimize contact balance for the asymmetric gripper.
Physical Validation: Comprehensive physical experiments comparing the proposed planners against a baseline (2D Bounding Box) using a UR5e robot arm, demonstrating significant improvements in lift success, shake survival, and force resistance.

4. Experimental Results

Experiments were conducted on 10 diverse objects (including difficult-to-grasp items like rock climbing holds and turbine housings) using a UR5e robot with a Zed Mini depth camera.

Metrics: Lift Success (%), Shake Survival (%), and Peak Slip Force (N).
Baseline Comparison (GET-2D-1.0 vs. 2D Bounding Box):
- Lift Success: Improved from 65% (Baseline) to 95% (+46%).
- Shake Survival: Improved from 60% to 90% (+50%).
- Force Resistance: Improved from 2.50 N to 3.95 N (+58%).
- Reasoning: The baseline failed to utilize holes in objects (leading to over-opening) and often failed to engage all three contact surfaces.
GET-3D-1.0 vs. GET-2D-1.0:
- Performance: GET-3D-1.0 achieved 100% lift success and 95% shake survival, slightly outperforming the 2D planner.
- Trade-off: GET-3D-1.0 is significantly slower, averaging 17 seconds per plan compared to 683 ms for GET-2D-1.0.
- Force Resistance: Interestingly, GET-2D-1.0 achieved a higher average slip force (3.95 N) than GET-3D-1.0 (3.20 N), suggesting the 2D planner's sampling strategy found more robust grasps in the specific test set.

5. Significance and Conclusion

This work bridges the gap between the novel hardware design of the GET gripper and practical robotic manipulation software.

Practicality: GET-2D-1.0 offers a near-real-time solution (sub-second) that significantly outperforms traditional bounding box approaches, making it suitable for dynamic industrial environments.
Robustness: The asymmetric geometry, when paired with these planners, provides superior torque resistance and stability compared to symmetric grippers.
Future Work: The authors note limitations regarding black objects (segmentation issues) and the computational cost of the 3D planner. Future directions include generating datasets for learning-based grasp planners to further accelerate the process.

In summary, the paper demonstrates that by tailoring grasp planning algorithms to the specific asymmetric geometry of the GET gripper, robots can achieve significantly higher success rates and stability in manipulating complex objects compared to standard symmetric approaches.