Metric 3D Visuomotor Learning
Low-data control, viewpoint robustness, and multi-view recovery
I built a calibrated RGB-D→world-frame point-cloud pipeline and a compact PointNet policy for closed-loop Panda reaching, then mapped both its advantages and its calibration failure boundaries.
Impact
- 95% closed-loop success with 50 demonstrations
- 96.7–100% success across 0–45° camera movement
- Two-view fusion restored strong-occlusion performance from 3.3% to 100% with no policy retraining
- Quantified collapse under depth and extrinsic calibration error
The result is scoped to contact-free reach with calibrated RGB-D and ground-truth segmentation; it is not presented as a universal 3D-over-RGB claim.