Are there any pose estimation models that work reliably in occlusion-heavy environments for grasp detection?

Yes — but with an important caveat: in heavy occlusion, no pose estimator is “reliably perfect” in a general sense. The best results usually come from models that are trained specifically for partial visibility, clutter, and grasp-like object poses, often combined with depth sensing or instance segmentation.

What tends to work best

For grasp detection in occluded scenes, these model families are commonly the most robust:

1. 6D object pose estimators with occlusion-aware training

These predict object orientation and position even when only part of the object is visible.

Examples:

• PVNet — keypoint voting makes it fairly robust to occlusion.
• DenseFusion — uses RGB-D and can handle partial views better than RGB-only methods.
• CosyPose — strong in cluttered scenes, especially with multi-view or RGB-D.
• FoundationPose — newer, strong zero-/few-shot style performance, often good under occlusion if depth is available.

Why they help for grasping:

• They infer the full object pose, which can be converted into grasp candidates or fed into a grasp planner.
• Depth data greatly improves robustness when surfaces are partially hidden.

2. Direct grasp detection models

Instead of estimating full object pose first, these predict grasp rectangles or 6D grasps directly from the scene.

Examples:

• GG-CNN — fast, works well on depth images, often used in clutter.
• Contact-GraspNet — predicts 6-DoF grasps directly from point clouds and is relatively strong in cluttered/occluded scenes.
• GraspNet-based models — the GraspNet benchmark specifically targets clutter/occlusion.
• AnyGrasp — designed for general 6D grasping in cluttered scenes and tends to be more practical than classic pose-first pipelines.

Why they help:

• They avoid needing a full clean pose estimate.
• They can propose grasps on visible geometry only, which is useful when objects are partially hidden.

3. Occlusion-trained keypoint/part-based methods

These are often more robust than dense regression under heavy occlusion because they rely on sparse visible evidence.

Examples:

• Keypoint voting approaches like PVNet
• Part-based 3D detectors
• Methods trained on synthetic occlusion augmentation

Practical recommendation

If your goal is grasp detection in occlusion-heavy environments, I’d usually recommend:

• Best practical choice with depth:
  Contact-GraspNet or AnyGrasp
• If you need object pose first:
  CosyPose, DenseFusion, or FoundationPose
• If you only have RGB:
  Occlusion robustness drops a lot; pose estimation becomes much less reliable. RGB-D is strongly preferred.

What improves reliability the most

Regardless of model, these usually matter as much as the architecture:

• Depth camera instead of RGB-only
• Synthetic occlusion augmentation during training
• Multi-view sensing if possible
• Instance segmentation before pose/grasp prediction
• Temporal tracking across frames
• Physical priors about object shape and grasp affordances

Bottom line

If you mean “works reliably in heavily occluded clutter for actual robot grasping”, the most practical answer is:

• Direct 6D grasp models on depth/point clouds are usually more reliable than classical pose estimators.
• If you need pose, use occlusion-robust RGB-D pose estimation like CosyPose, DenseFusion, or FoundationPose, ideally combined with a grasp planner.

If you want, I can also give you:

1. a short list of the best current open-source models, or
2. a recommendation based on your sensor setup (RGB, RGB-D, stereo, point cloud).