Project Retrospective: Grasping in Clutter

Background

Bin picking in a cluttered scene with partial observability.

Goals

• Achieve 80 percent success rate on 20 objects.
• End-to-end pipeline with perception and grasping.

Approach

• Train grasp proposals in simulation.
• Fine-tune on real data with domain randomization.
• Integrate with a motion planner and force feedback.

Results

• Reached 76 percent in the lab with stable runtime.

Lessons Learned

• Dataset bias dominated early failures.
• Calibration drift required weekly checks.

Next Actions

• Expand object set and improve camera poses.