Simulated bin picking: rigid versus deformable

The problem

Every gripper vendor publishes success rates on rigid boxes. Almost nobody publishes the same number for a crumpled poly mailer. The deformable case is exactly the one that breaks deployments — so it is the one worth measuring.

This build measures the gap myself, in simulation, with a suction-cup model run against both object classes under matched conditions.

The build

A PyBullet scene: a bin, a suction end-effector, and two object sets — rigid boxes and soft-body mailers. Each trial drops a mix into the bin, runs a top-down pick policy, and logs whether the grasp held to the place pose.

The point isn't a fancy policy. It's a clean, repeatable measurement of how much pick success degrades when the only thing that changes is the deformability of the item.

What broke

Every project on this site has this section, because the whole point of the site is that demos lie by omission. Here is what actually went wrong:

• Soft-body simulation tanked the physics timestep. Stable rigid-body settings made the deformable mesh explode; stable deformable settings made the rigid run pointlessly slow.
• The first deformable mesh clipped straight through the bin floor for two days before I fixed the collision margins.
• Suction contact on a slumping surface is genuinely ambiguous — "did it grab" needed a real definition, not a vibe.

Results

Honest numbers when there are numbers. The rig runs end to end; the measurement pass is in progress. No headline figure until the protocol is locked, because a number I can't reproduce is worse than no number.

Next rev

Rev C: vary mailer fill levels; add a parallel-jaw gripper model for comparison; publish the full trial log alongside the writeup.