# Does Bimanual Grasping Actually Work Without a Full 3D Scan?

A new paper from researchers at IIIT Hyderabad directly attacks one of the most persistent assumptions in dual-arm manipulation research: that a robot needs a complete point cloud of an object before it can grasp it with both hands. PartialBiGrasp, authored by Ayush Kaura, Vignesh Vembar, Md Faizal Karim, Keshab Patra, and K Madhava Krishna, proposes a grasp generation framework that works directly from partial point cloud observations — the kind of incomplete, noisy depth data that a real robot actually sees in deployment.

The core insight is deceptively simple but practically significant. Large, heavy, geometrically complex objects — exactly the class of objects that *require* bimanual handling — tend to have sparse graspable regions dictated by local geometric properties: thickness, edge structure, and gripper clearance. When a robot only sees one side of an object, those graspable regions on occluded surfaces are precisely what's missing. Prior bimanual methods assumed away this problem by requiring full point clouds. PartialBiGrasp doesn't.

The system uses convolutional occupancy networks to learn geometric features implicitly, enabling the model to reason locally about graspability, collision-free contact regions, and object thickness without requiring a complete reconstruction. Force-closure compliant grasp pairs are then refined via sampling-based optimization to correct for ambiguity introduced by the incomplete geometry.

Validation covers analytical force-closure metrics, large-scale simulation experiments, and real-world robot evaluations on noisy partial point clouds of novel objects.

## Why the "Full Point Cloud" Assumption Has Always Been a Problem

The field of [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) has long operated with a quiet inconsistency: lab demonstrations almost invariably use objects that have been carefully scanned, reconstructed, or placed in known configurations, while deployment environments — factory floors, warehouses, homes — offer none of those guarantees.

For single-arm grasping, this gap has been partially addressed. Grasp pose detection methods that operate on single-view depth images are now relatively mature, and [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) to novel objects has improved considerably with learned feature representations. Bimanual grasping has lagged because the coordination constraint is dramatically harder: two end-effectors must simultaneously achieve force closure, avoid colliding with each other, and account for object geometry that may be invisible to the sensor.

The standard workaround — require a full scan — is operationally impractical. In real manipulation scenarios, a robot approaches an object from a direction, sees part of it, and needs to act. A full 360-degree scan either requires the robot to move around the object (time-consuming and not always feasible in cluttered environments) or relies on a sensor configuration that most deployed humanoid platforms don't have.

The authors' use of convolutional occupancy networks here is worth noting. Rather than explicitly predicting the missing geometry through shape completion — an approach that accumulates reconstruction error before grasp planning even begins — PartialBiGrasp learns to reason about graspability *implicitly* from the partial observation. The model infers what the local geometry probably supports without needing to reconstruct what it can't see. That's a meaningfully different architectural choice, and one that sidesteps a compounding failure mode.

## Sampling-Based Optimization as an Ambiguity Corrector

The second architectural element worth examining is the sampling-based optimization stage used to refine generated grasp pairs. Partial observations are inherently ambiguous — the same visible surface could be consistent with multiple hidden geometries, and a grasp that looks viable on the observed portion may collide with or fail to engage the occluded portion.

Rather than attempting to resolve this ambiguity deterministically, the authors treat it as an optimization problem: generate candidate grasp pairs, then use sampling to search for configurations that are robust across the plausible range of hidden geometries. This is a sensible engineering choice. It acknowledges that uncertainty exists and manages it explicitly, rather than hoping a single forward pass produces a correct answer.

The tradeoff is inference time. Sampling-based refinement is computationally heavier than a single-pass prediction, which matters for real-time control loops on humanoid platforms. The paper validates on real-world hardware, which suggests the compute cost is at least tractable, but the authors don't claim real-time performance in the abstract — engineers considering integration should probe that directly.

## What This Means for Humanoid Bimanual Manipulation

The practical implications map directly onto where humanoid robotics is heading. Nearly every major humanoid platform in active deployment or late-stage development — from [Figure AI](https://humanoidintel.ai/companies/figure-ai)'s Figure 02 to [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics)' Digit — is designed around bimanual manipulation as a core capability. Moving heavy boxes, handling large panels, assembling components that require two hands: these are the tasks that justify the form factor.

The manipulation stack for most of these platforms still relies heavily on known object models or carefully structured environments. As deployments move from controlled pilot sites into genuine unstructured settings, the partial-observation problem will become the central bottleneck. Research like PartialBiGrasp addresses that bottleneck at the grasp planning layer.

The connection to [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) is also worth flagging. Bimanual grasping doesn't happen in isolation on a humanoid — the robot is also managing balance, foot placement, and torso posture while its arms coordinate. A grasp planner that produces physically stable, force-closure compliant grasps from partial observations is a more composable module for whole-body controllers than one that requires a pre-processing pipeline of full reconstruction. Cleaner interfaces between perception, grasp planning, and whole-body control are an underappreciated engineering priority.

## Skeptical Notes

A few caveats merit attention before this work gets adopted into deployment pipelines.

**Generalization scope:** The paper validates on "novel objects" in real-world evaluation, but the definition of novelty matters enormously. Objects that are geometrically similar to training data are not a meaningful test of zero-shot generalization. The abstract doesn't detail the object diversity of the test set — that information is in the full paper and should be reviewed carefully.

**Gripper specificity:** Local graspability depends heavily on gripper geometry — clearance, jaw width, approach angle constraints. The framework reasons about gripper clearance as part of local geometry, but it's not clear from the abstract how tightly the model is coupled to the specific gripper used in training. Humanoid platforms use a wide variety of end-effector designs, from parallel-jaw grippers to multi-fingered hands, and transfer between them is rarely free.

**Force-closure as a proxy:** Analytical force-closure metrics are the field's standard evaluation, but they measure a necessary, not sufficient, condition for successful grasping. Real-world grasp success also depends on surface friction, compliance, and dynamic effects that force-closure analysis doesn't capture. The real-world evaluations are the more meaningful signal here.

---

## Key Takeaways

- **PartialBiGrasp** (arXiv:2608.19188) generates dual-arm grasps directly from partial point clouds, removing the full-scan assumption that breaks bimanual grasping in real deployment.
- The system uses **convolutional occupancy networks** to implicitly learn local geometric features — thickness, edge structure, gripper clearance — without requiring explicit shape completion.
- **Sampling-based optimization** refines grasp pairs to correct for ambiguity caused by incomplete geometry observation.
- Validation includes analytical force-closure metrics, large-scale simulation, and real-world robot experiments on noisy partial point clouds of novel objects.
- For humanoid platforms targeting unstructured bimanual manipulation, partial-observation grasp planning is becoming the critical capability gap — and this paper offers one credible approach to closing it.
- Practitioners should scrutinize object diversity in the test set and gripper transferability before integrating this into production pipelines.

---

## Frequently Asked Questions

**What is PartialBiGrasp?**
PartialBiGrasp is a dual-arm grasp generation framework that produces force-closure compliant grasp pairs for both robot arms using only partial point cloud observations of an object, without requiring a complete 3D scan. It was developed by Ayush Kaura, Vignesh Vembar, Md Faizal Karim, Keshab Patra, and K Madhava Krishna.

**Why is bimanual grasping harder than single-arm grasping?**
Bimanual grasping requires two end-effectors to simultaneously achieve stable contact, avoid mutual collision, and account for the full geometry of often large and heavy objects. When sensor data is incomplete — as it always is in real deployments — coordinating both arms without knowing the full object shape is a substantially harder planning problem than single-arm grasping.

**What are convolutional occupancy networks and why do they matter here?**
Convolutional occupancy networks are neural architectures that learn to represent 3D geometry implicitly, predicting whether a given point in space is occupied or free. In PartialBiGrasp, they enable the model to reason about local geometric properties — like object thickness and gripper clearance — from partial observations, without needing to first reconstruct the complete object shape.

**How does this research apply to humanoid robots specifically?**
Humanoid platforms are increasingly targeting bimanual manipulation of large, heavy objects as a core commercial use case. Most current manipulation stacks assume structured environments or known object models. PartialBiGrasp's approach to handling partial observations makes it more compatible with the unstructured, real-world settings where humanoid deployments are heading.

**What are the main limitations of this approach?**
Key open questions include how well the model generalizes to object geometries far outside its training distribution, how tightly it is coupled to the specific gripper used during training, and whether the sampling-based refinement stage is fast enough for real-time control on deployed humanoid platforms.