What's the best robot perception model for 3D perception in an occlusion-heavy warehouse environment?

There isn’t a single universally “best” model, but for an occlusion-heavy warehouse the strongest choice is usually:

Best practical choice

A multi-modal 3D perception stack built around LiDAR + RGB, with 3D detection/segmentation models such as:

• BEVFormer / BEVFusion for robust bird’s-eye-view scene understanding
• CenterPoint for 3D object detection from LiDAR
• PV-RCNN++ if you want a very strong LiDAR-only detector
• Mask3D or SparseConv/3D U-Net-style models for 3D instance/semantic segmentation

If you want one recommendation

For a warehouse with lots of occlusion, I’d pick:

BEVFusion + CenterPoint-style LiDAR detection

Why:

• LiDAR handles occlusion better than camera-only methods
• BEV representations are excellent for navigation, mapping, and tracking
• Fusion with cameras helps with object classification and fine-grained cues
• Works well for forklifts, pallets, racks, people, cartons in cluttered spaces

Why not camera-only?

Camera-only 3D perception models like:

• Monocular 3D detection
• DETR3D
• BEVFormer with only RGB

can struggle in warehouses because:

• objects are frequently partially hidden
• lighting can vary a lot
• depth from images alone is ambiguous

If your goal is navigation rather than detection

If you care more about mapping free space and traversability, use:

• 3D semantic occupancy networks
• BEV-based occupancy prediction
• Voxel/TSDF mapping + learned semantics

Good options:

• BEVFusion / BEVFormer occupancy heads
• MonoScene-style occupancy methods
• SurroundOcc for dense occupancy estimation

Best model depends on the task

• 3D object detection: CenterPoint, PV-RCNN++
• Multi-sensor perception: BEVFusion, BEVFormer
• Segmentation/instance understanding: Mask3D
• Occupancy/free-space prediction: SurroundOcc, occupancy networks

My short answer

For an occlusion-heavy warehouse, the best overall approach is LiDAR + camera fusion using a BEV-based model, especially BEVFusion, paired with a strong LiDAR detector like CenterPoint.

If you want, I can also give you:

1. a ranked list of the top 5 models,
2. a recommended model for ROS2 deployment, or
3. a comparison table for warehouse robotics.