💻 computer science

BiDexGrasp: Coordinated Bimanual Dexterous Grasps across Object Geometries and Sizes

This paper introduces BiDexGrasp, a comprehensive framework that combines a novel two-stage synthesis pipeline for constructing a large-scale bimanual dexterous grasp dataset with 9.7 million annotations and a geometry-size-adaptive generative model to enable coordinated, high-quality grasping on unseen objects.

Original authors: Mu Lin, Yi-Lin Wei, Jiaxuan Chen, Yuhao Lin, Shuoyu Chen, Jiangran Lyu, Jiayi Chen, Yansong Tang, He Wang, Wei-Shi Zheng

Published 2026-04-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mu Lin, Yi-Lin Wei, Jiaxuan Chen, Yuhao Lin, Shuoyu Chen, Jiangran Lyu, Jiayi Chen, Yansong Tang, He Wang, Wei-Shi Zheng

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 you are trying to pick up a giant, awkwardly shaped watermelon with two hands. If you try to grab it with just one hand, it might slip, or you might not be able to reach the right spot. But with two hands, you can stabilize it, lift it, and move it safely. This is the challenge of bimanual dexterous grasping in robotics: teaching robots to use two "hands" to pick up all kinds of weird, big, and small objects, just like humans do.

The paper introduces BiDexGrasp, a new system that solves two major problems holding robots back: lack of practice data and clumsy AI models.

Here is the breakdown using simple analogies:

1. The Problem: The Robot's "Blank Canvas"

Before this paper, robots trying to learn how to use two hands were like art students trying to paint a masterpiece without ever seeing a reference photo.

The Data Gap: Existing datasets were tiny and only showed robots picking up small, simple objects. They didn't have enough examples of how to handle big, weird, or heavy things with two hands.
The "Search" Nightmare: Trying to figure out where to put two hands on a 3D object is like trying to find a specific needle in a haystack, but the haystack is the size of a football stadium, and you have to find two needles at the same time that work together. Old methods were too slow and often gave up.

2. The Solution Part 1: The "Super-Producer" (The Dataset)

The authors first built a massive library of "how-to" examples. They didn't just guess; they built a synthesis pipeline (a factory) to create 9.7 million perfect examples of two-handed grasps.

The "Smart Search" (Region-Based Initialization): Instead of randomly guessing where to put the hands (which is like throwing darts blindfolded), their system first looks at the object and says, "Okay, these two spots look like they could hold the object well." It narrows down the search to the best neighborhoods.
The "Decoupled" Strategy: Imagine two people trying to lift a heavy couch. If they try to coordinate every single muscle movement at once, they might trip. Instead, BiDexGrasp tells the left hand, "You focus on getting a good grip here," and the right hand, "You focus on getting a good grip there." Once both have a solid grip, they lock together. This makes the math much faster and the result much stronger.
The Result: They created a dataset with 6,351 different objects ranging from small (30cm) to huge (80cm). This is like giving the robot a library of every possible object it might ever encounter.

3. The Solution Part 2: The "Smart Brain" (The AI Model)

Once the robot has this massive library of examples, they trained a new AI model (BiDexGrasp) to learn from it. This model has two special superpowers:

The "Team Captain" (Bimanual Coordination): Most AI models just tell the left hand what to do and the right hand what to do separately. This new model acts like a team captain. It looks at the object and says, "Left hand, you take the top; Right hand, you take the bottom." It ensures the two hands work together to pick a stable spot, avoiding collisions between the hands.
The "Shape-Shifter" (Geometry-Size Adaptation): Robots usually struggle when an object is bigger or smaller than what they trained on. This model is like a chameleon. It doesn't just memorize the size of the object; it learns the shape and structure. Whether the object is a tiny cup or a giant box, the model adjusts its "grip anchor" (the starting point for the grasp) to fit perfectly.

4. The Results: From Simulation to Reality

The team tested this in two ways:

In the Computer (Simulation): They ran millions of virtual tests. Their method was 2.8 times more successful at grabbing objects than previous methods and 30 times faster at generating the data.
In the Real World: They put the AI on real robots with real hands (like the Shadow Hand and Inspire Hand). They tested it on objects the robot had never seen before.
- The Outcome: The robot successfully grabbed tricky objects with a very high success rate (up to 100% in some tests). It could lift, hold, and move objects without dropping them.

The Big Picture Analogy

Think of the old way of teaching robots as trying to teach a child to juggle by only showing them pictures of juggling apples. When you give them a bowling ball, they fail.

BiDexGrasp is like:

Hiring a team of experts to film millions of videos of people juggling everything from ping-pong balls to watermelons (The Dataset).
Teaching the robot a new way to watch those videos, where it learns to coordinate its two hands like a dance partner and adapts its moves based on how heavy or big the object is (The Framework).

In short: BiDexGrasp gives robots the "muscle memory" and the "brainpower" to pick up almost anything with two hands, making them much more capable of helping us in the real world.

1. Problem Definition

Bimanual dexterous grasping is essential for robots to manipulate large, heavy, or complex objects that exceed the capabilities of a single hand. However, progress in this field is hindered by two main challenges:

Data Scarcity: Existing datasets are limited in scale, diversity (object sizes and geometries), and quality. Most current methods rely on extending single-hand pipelines, which fail to address the high-dimensional search space and coordination requirements of two hands.
Optimization Complexity: Directly optimizing bimanual grasps leads to a quadratic growth in initialization candidates and a high-dimensional force-closure optimization space. This results in low efficiency, poor convergence, and suboptimal solutions where one hand dominates the grasp while the other contributes little.
Generalization: Existing models struggle to generalize across objects with vastly different sizes (30cm to 80cm) and geometries.

2. Methodology

The authors propose BiDexGrasp, a comprehensive solution comprising a large-scale data synthesis pipeline and a novel learning-based generation framework.

A. BiDexGrasp Dataset Synthesis Pipeline

To overcome data limitations, the authors developed a three-stage pipeline to generate 9.7 million high-quality bimanual grasp annotations for 6,351 diverse objects (sizes 30–80 cm).

Region-Constrained Initialization:
- Instead of uniform sampling (which is inefficient), the pipeline uses Grasp Wrench Space (GWS) analysis to identify stable grasping regions.
- It selects top candidate region pairs based on stability scores against external disturbances, significantly reducing the search space.
Decoupled Force-Closure Optimization:
- Traditional bimanual optimization couples the hands, leading to imbalanced solutions.
- The authors propose a decoupled strategy that separates the optimization energy into single-hand terms ( $Q(G_{left})$ and $Q(G_{right})$ ). This ensures both hands achieve high-quality local contacts independently while minimizing a global loss function that includes collision avoidance and region consistency.
Kinematics-Aware Filtering:
- The pipeline performs Inverse Kinematics (IK) and collision checking to ensure the generated grasps are physically reachable by robot arms and free of self-collisions.

Performance: This pipeline achieves 2.8× higher success rates and 30× faster synthesis speeds compared to previous bimanual methods.

B. BiDexGrasp Generation Framework

A data-driven generative model is trained on the synthesized dataset to predict grasps for unseen objects. It addresses two key challenges: coordination and adaptation.

Bimanual Coordination Module (BCM):
- Predicts a pair of coordinated grasp views (approach directions) for the two hands.
- It uses cross-attention to extract view features and models the interaction between the primary view and secondary views to ensure global stability and avoid inter-hand collisions.
Geometry-Size-Adaptive Strategy:
- Scale-Adaptive Grasp Anchor: Instead of predicting absolute poses, the model predicts relative poses ( $\Delta t, \Delta r$ ) relative to a dynamic grasp anchor derived from the object's minimum bounding sphere. This reduces the action space and improves scale invariance.
- Geometry-Adaptive Features: The model extracts local geometric features around the grasp anchors (using KNN) rather than relying solely on global features, allowing it to focus on relevant surface structures.
Relative Dexterous Grasp Diffusion:
- Utilizes a Denoising Diffusion Probabilistic Model (DDPM) to generate the relative wrist poses and finger joint configurations conditioned on the object point cloud and the adaptive anchors.
- The training loss includes physics-based penalties for hand-object penetration and self-penetration.

3. Key Contributions

Large-Scale Dataset: Construction of the largest bimanual dexterous grasp dataset to date, containing 6,351 objects and 9.7 million annotated grasps, covering a wide range of sizes (30–80 cm) and geometries.
Novel Synthesis Pipeline: Introduction of a region-constrained initialization and decoupled force-closure optimization strategy that drastically improves data generation efficiency and quality.
Adaptive Generation Framework: A novel deep learning framework that explicitly models bimanual coordination and employs geometry-size-adaptive mechanisms (relative pose prediction and local feature extraction) to generalize to unseen objects.
Real-World Validation: Successful deployment on physical robots (BrainCo/Inspire hands with Piper/Unitree arms), demonstrating the transferability of the simulation-trained model to the real world.

4. Experimental Results

Data Synthesis: The pipeline achieved a 77.11% lift success rate (SUC-A) in tabletop scenarios, significantly outperforming the baseline BimanGrasp (26.8%). It also demonstrated a 40× speedup in synthesis speed.
Generation Framework:
- On the DexGraspNet test set, the framework achieved a 66.78% lift success rate, outperforming state-of-the-art methods like DGTR (41.45%) and SceneDiffuser (15.17%).
- It maintained the lowest penetration depth (PD) and self-penetration depth (SPD), indicating physically feasible and stable grasps.
Real-World Experiments: Tested on unseen objects with different robot hands, the framework achieved success rates ranging from 50% to 100% depending on object complexity, validating its robustness in real-world settings.
Ablation Studies: Confirmed that both the Bimanual Coordination Module and the Geometry-Size-Adaptive strategy are critical for performance, with the full model outperforming ablated versions by a significant margin.

5. Significance

BiDexGrasp represents a major step forward in robotic manipulation by addressing the "data bottleneck" in bimanual dexterous grasping. By providing a scalable method to generate high-quality data and a robust framework to learn from it, the work enables robots to handle a much wider variety of real-world objects. The decoupled optimization strategy and adaptive generation approach offer new paradigms for solving high-dimensional coordination problems in robotics, paving the way for more capable and generalizable dual-arm robotic systems.