Scenario Lab
Scenario generator for VLA imitation learning

Edge-case household scenarios for VLA training data.

A scenario generator for robotics teleoperation operators. The diversity of your training demonstrations is what shapes the model — this tool helps surface the unusual, occluded, partial-visibility cases that fresh operators don't think to run.

Recent scenarios

View all →

Deflated Kiddie Pool Tangled in Overgrown Lantana After Last Night's Wind

Outdoor D5 W5

You are positioned at the edge of a back patio, facing a small suburban yard in mid-morning light — overcast sky diffusing shadows so depth...

#outdoor #malleable_objects #occlusion #deformable_pvc

Night Owl Laundry: Cat Nesting in Damp Sheets

Laundry Room D4 W4

The laundry room is lit only by the faint blue glow of the washer's cycle-complete indicator light and a thin strip of amber light bleeding...

#low_light #animal_in_scene #deformable_objects #multi-step_planning

The Half-Dismantled Bike and the Mystery Box Blocking the Drain

Garage D5 W4

You are in a cluttered two-car garage that currently holds zero cars. Fluorescent overhead lights flicker faintly on one side; the other hal...

#garage #multi_step #ambiguous_goal #drain_clearance

Sprinkler Chaos: Retrieve Soaked Backpack Before School Bus Arrives

Outdoor D5 W4

You are positioned on a concrete back patio, looking out across a partially sunlit backyard at approximately 7:40 AM. The automatic irrigati...

#outdoor #kids #time-pressure #deformable-object

Leaking Coolant Drip Tray Before the Car Returns

Garage D5 W3

You are facing the interior of a single-car garage at approximately 6:45 PM. The main overhead fluorescent tube on the left is flickering at...

#garage #time_pressure #occlusion #fluid_handling

Deflated Mylar Birthday Balloon Tangle Under the Couch

Living Room D4 W5

The operator's view opens at chest height, facing a cluttered living room in late-afternoon light — sun is low through half-closed horizonta...

#malleable_objects #deformable #mylar #ribbon_entanglement

Why this exists

01

Imitation learning data is the bottleneck.

VLA models are constrained by what they've seen demonstrated. More diverse demonstrations beat more parameters.

02

Clean lab data underrepresents real homes.

A robot trained only on staged scenes fails when it meets a cat in the laundry basket. This tool nudges operators toward the messier cases.

03

Structured narration improves grounding.

Consistent action and spatial vocabulary in narration scripts gives the model cleaner language signal to anchor visual concepts to.