Simulation-ready 3D assets for warehouse robotics
Pickle Robot trains trailer-unloading robots on synthetic data generated in NVIDIA Isaac Sim. Every box, trailer, and strip of tape in that simulator had to look right and behave right, because an asset that only looks right corrupts the data the robot learns from.
The result: a full asset library delivered to spec and running in Pickle Robot's training pipeline, with three hidden defects already in production caught and fixed along the way.
Robotics
Synthetic Data
NVIDIA Isaac Sim
OpenUSD
3D Production
3
BOX
SIZES
TRAILER &
VEHICLE TYPES
HIDDEN DEFECTS
CAUGHT
THE CONSTRAINT
Warehouse robots learn in simulation before they touch a real trailer. The simulator generates thousands of labeled scenes of boxes, tape, and trailers under varied lighting and positions, and that data is only as good as the 3D assets inside it.
Standard library assets fail this job. Wrong scale breaks physics. Missing collision means the robot trains on objects it can never touch. Broken materials produce artifacts that contaminate the dataset. An asset can pass any visual review and still poison the training data.
The brief: deliver Pickle Robot's full asset inventory to their Isaac Sim spec, exactly, at production pace.
WHAT I DID
The spec became the stakeholder. Every rule got read for the reason behind it: metric scale and Z-up because the physics engine assumes meters, collision present but invisible because the robot touches what the camera never sees, consistent origins so scene assembly never special-cases an asset, and PBR texture sets so materials read correctly under the simulator's lighting.
Repetitive spec work moved into an AI-assisted pipeline, with Claude writing Python for Blender and headless USD tooling. Human attention went where automation can't judge: verification, and the hand-painted wear that makes variants read as real.
THE TURNING POINT
A formatting request became an audit
Midway through, Pickle Robot asked for per-surface texturing so their simulator could randomize floors, walls, and ceilings independently. Doing it right meant checking what the delivered files actually contained, and that check found three defects already running in production.
KEY FINDINGS
Five texture files named for five surfaces were only two images. All five interior materials had their normal map and ORM inputs swapped, so floors and ceilings rendered black under direct light. The box truck's cargo walls had zero thickness, so the renderer could not decide which face to draw. Each one passed visual review.
The fix went past the request: per-surface texture sets, corrected shader inputs, and the box truck's interior skin moved two millimeters inward. Verification became part of every delivery after that, through pixel comparison, material raycast checks, measured test renders, and USD compliance checks.
KEY DECISIONS
REAL GEOMETRY FOR PEELING TAPE
The trade-off. The spec allowed flat textured planes, which are fast to produce and render fine. They train badly, because the perception model needs true depth cues. Peeling tape was built as displaced 3D geometry instead: slower to produce, correct for the purpose.
LICENSED VEHICLES, REBUILT TO SPEC
The trade-off. Modeling a 53-foot trailer from scratch would have cost days with no gain in training value. Trailers and vans started from licensed commercial models and were rebuilt to spec: rescaled to metric, re-origined, collision added, materials renamed, textures re-pathed. Effort went where the simulator benefits.
OUTCOMES
A simulation asset library delivered to spec and running in Pickle Robot's training pipeline.
Inventory Delivered To Spec Boxes in four sizes with damage, flap, and tape states, plus trailers and vehicles, an environment, and decals, all shipped as USD into the active simulation pipeline.
Hidden Defects Fixed
Three defects already running in production, duplicate textures, swapped shader inputs, and coplanar walls, found through measurement and fixed.
Interiors Ready For Randomization
Per-surface texture sets let the simulator vary floors, walls, and ceilings independently for training variety.
Verification Habit, Built Into The Pipeline
Pixel comparison, material raycast checks, measured test renders, and USD compliance checks on every export before delivery.
What I'd do differently
Build verification checks on day one. The swapped inputs and duplicate textures surfaced late because automated checks came partway through. Asking for per-surface texturing up front would have avoided converting atlas interiors after delivery, and testing in the client's environment early would have caught hidden colliders dropping when files opened directly in Isaac Sim.
Next Steps
The verification checks built during this project could run inside the client's pipeline, so every new asset gets tested on arrival instead of after a defect surfaces. The library could also extend to more trailer types and box conditions as the robot meets new scenarios in the field.