VulcanVoxel: 3D Affordance Model Boosts Blade Insertion in Cluttered Bins
30 ms RGB-to-voxel inference achieves 89% coverage versus 71% for pose-based baselines.
VulcanVoxel rethinks robotic stowing by treating blade affordances as spatial objects rather than SE(3) pose distributions. Instead of predicting where a blade should go, the model reconstructs blade occupancy conditioned on scene geometry, computing feasibility locally at each voxel. This spatial approach recovers multi-modal predictions from unimodal execution data, bypassing the structural limitations of pose-based methods.
Trained on 10,000 real warehouse stow episodes without human annotation, VulcanVoxel achieves top-5 coverage of 0.89 versus 0.71 for the best pose-based baseline. A distilled student model further accelerates RGB-to-voxel inference to 30 ms (down from 1.4 s), enabling real-time deployment. The team has released a dataset of blade insertion cycles with RGB-D observations and pose trajectories.
- VulcanVoxel uses a masked autoencoder over 3D occupancy fields to predict feasible blade insertion spaces.
- Achieves 0.89 top-5 coverage on 10,000 real warehouse episodes, beating 0.71 pose-based baseline.
- Distilled student model runs RGB-to-voxel inference in 30 ms (vs 1.4 s for voxel-to-voxel).
Why It Matters
Enables faster, more accurate robotic stowing in cluttered warehouses, reducing errors and increasing throughput without human labels.