Everyday tasks such as carrying boxes, lifting trays, and moving chairs naturally require coordinated interaction between two robotic arms to support, stabilize, and manipulate them. By distributing contacts between both manipulators, bimanual grasping improves object grasp stability and load balancing, enabling robust manipulation of bulky objects that would otherwise induce unstable interactions for a single arm. Despite its importance, relatively few works have explored learning-based bimanual grasp generation, particularly under partial observation settings.
Introduction
To address these challenges, PartialBiGrasp generates stable dual-arm grasps directly from a single RGB-D observation. Rather than relying on complete object geometry or explicit shape reconstruction, the method reasons about hidden local geometry to predict force-balanced and collision-free grasp pairs for large objects with sparse graspable regions.
PartialBiGrasp is, to the best of the authors’ knowledge, the first framework for bimanual grasp generation on single-view partial point clouds. It decomposes the task into single-grasp generation, force-closure-based grasp pairing, and local-geometry-conditioned refinement, enabling efficient generation of stable dual-arm grasps under partial observations.
The method introduces an occupancy-conditioned geometric reasoning pipeline that leverages convolutional occupancy networks and grasp-aligned local features to implicitly infer graspability, local object geometry, collision-free contact regions, and gripper feasibility, together with a sampling-based refinement procedure to improve grasp quality.
The approach is evaluated through analytical force-closure validation, large-scale simulation experiments, and real-world robot deployment, demonstrating robust dual-arm grasp generation and successful manipulation across diverse large-object scenarios.
Dual-Arm Grasp Synthesis
Using the implicit geometry representation described above, dual-arm grasps are generated in two stages. First, feasible single-arm grasp candidates are predicted from the observed object geometry. Candidate grasp pairs are then evaluated using a learned force-closure critic.
Experimental Setup
All models are trained on the DG16M dataset. Training is performed in three stages.
First, the Global Geometry Encoder and the single-grasp synthesis heads are jointly trained using occupancy supervision, grasp-pose supervision, and graspability-region supervision derived from valid grasp contacts in DG16M.
Next, the Dual-Grasp Critic is trained to classify whether candidate grasp pairs satisfy force-closure constraints.
Finally, the Local Geometry Encoder is trained using local occupancy supervision to enable occupancy-guided grasp refinement under partial observations.
The framework is tested on an unseen split from DG16M and on a collection of real-world objects captured using a RealSense D455, for which digital twins were created. All experiments are performed on a single NVIDIA RTX 3090 GPU.
Limitations
The framework does not explicitly model grasp reachability or motion feasibility for a given embodiment during grasp generation. As a result, some predicted grasp pairs may be infeasible on dual-arm systems because of kinematic constraints or arm collisions, requiring a motion planner to validate and execute grasps for deployment.
Dataset
The method is trained and evaluated using the DG16M dataset, which contains 4,142 large-object meshes from ShapeNet annotated with approximately 2,000 positive and negative dual-arm grasp pairs per object generated using analytical force-closure evaluation.
Three hundred objects are reserved for testing, and the remaining objects are used for training. For each object, single-view partial observations are generated by rendering RGB-D images from randomly sampled viewpoints and converting the visible depth measurements masked by the object into partial point clouds. All quantitative results are reported on this unseen test split.
To train the single-grasp generation module, unique valid single-arm grasps are extracted from the grasp pairs provided in DG16M. The resulting grasp poses and contact locations are used as supervision for grasp-pose regression.
Graspability labels are derived from these contact locations. Positive samples correspond to valid single-grasp contacts, while negative samples are drawn from regions assigned a zero graspability score by an annotated graspability map. The map is constructed by propagating graspability scores from valid single-grasp contact locations in DG16M using a distance-based weighting function over the object surface.
Baseline Details
ContactGraspNet is first used to generate single-arm grasp candidates from the observed partial point cloud. The high-scoring grasp candidates are filtered according to their predicted grasp quality. Farthest-point sampling then obtains a spatially diverse set of grasp candidates together with their associated point features.
Two pairing strategies are evaluated.
Classifier-Based Pairing
A binary grasp-pair classifier is trained using DG16M force-closure annotations. For each visible observation, ground-truth grasp pairs are matched to the nearest visible ContactGraspNet grasp candidates. The corresponding point features are concatenated and used to predict whether the pair satisfies force-closure constraints.
During inference, the classifier scores all candidate grasp pairs, and the highest-scoring candidate pairs are retained.
Results
The evaluation uses Force Closure (FC%), Grasp Success Rate (GS%), and Pair Collision Rate (PC%).
- FC% measures the average percentage of predicted grasp pairs that satisfy analytical force-closure constraints.
- GS% measures the average percentage of predicted grasp pairs that successfully lift and hold the object in simulation.
- PC% measures the average percentage of predicted grasp pairs that collide with the object geometry.
Higher FC% and GS% indicate better grasp-pair quality, while lower PC% indicates better geometric feasibility.
Coverage measures how well the set of force-closure-valid grasp pairs spans the graspable regions of the object.
Coverage is based on the number of positive graspability points enclosed by the volume between the two grippers across all force-closure-valid grasp pairs, divided by the total number of positive graspability points.
Real-World Demonstration
All real-world experiments are performed on a heterogeneous dual-arm platform consisting of an xArm7 and an xArm6 Lite equipped with parallel-jaw grippers. A RealSense D455 RGB-D camera is mounted above the workspace and provides the single-view RGB-D observations used by the grasping pipeline.
Camera calibration is performed using an ArUco marker board placed within the workspace. The generated dual-arm grasps are pruned to retain only collision-free and kinematically reachable grasp pairs before execution on the robot platform.
The resulting grasps are then executed to evaluate sim-to-real transfer under partial observations and previously unseen object geometries.
Frequently Asked Questions
What problem does PartialBiGrasp address? It generates stable dual-arm grasps for large objects from a single-view partial RGB-D observation.
How does the method handle hidden object geometry? It uses occupancy-conditioned geometric reasoning and local geometry features to implicitly infer graspability, contact regions, and collision-free grasp refinements.
What dataset is used for training and evaluation? The method uses DG16M, which contains 4,142 large-object meshes with approximately 2,000 positive and negative dual-arm grasp pairs per object.
What is a limitation of the framework? It does not explicitly model embodiment-specific reachability or motion feasibility, so a motion planner is required to validate and execute some predicted grasp pairs.