RobotFlyWheel-Soft
RFW-S: Everyday Deformable Objects
with Independent Simulation Verification
Everyday variety. Executable physics.
Evidence beyond the fitted interaction.
Read the abstract
Everyday deformable objects require both broad semantic coverage and physical configurations that can be tested under interaction. We introduce RobotFlyWheel-Soft (RFW-S), a dataset of 100,000 generated assets spanning 485 everyday categories across soft solids, cloth sheets, shells, garments, and bags. Construction connects reference appearances and geometry to structure-aware representations, executable simulator parameters, calibration feedback, and independent testing of the selected configuration. The intersection of physical verification and runtime checks retains 83.4% of the pool and all categories. In an expanded bag experiment, an independent stimulus rejects 3.87% of candidate evaluations that satisfy calibration, demonstrating the additional screening provided by testing beyond the fitted interaction. Garment comparisons characterize the effects of combined material and structural settings; a grasping application separates task completion from a deformation-derived integrity criterion. RFW-S associates semantic coverage with explicit configuration and behavioral evidence, supporting reproducible deformable-object studies within the documented solver and interaction scope.
Everyday diversity · Executable configurations
Explore RFW-S-100K
Food, fabrics, cushions, clothing, and carryalls.
Different forms. Different ways to move.
No matching assets
Try another category or show all families.
Click an object to explore its available views and interactions.
Appearance previews are normalized for visibility; individual verification status is not inferred.
Beyond one example
Same category. Different character.
Six categories, six instances each.
Explore variation in shape, color, and appearance.
Objects in action
Drop. Drape. Make contact.
Fifteen objects, in context.
Select an action to see its recorded response.
Illustrative simulations with recorded execution. These scenes demonstrate use contexts; they are not measurements of real-world objects.
Structure makes it executable
From appearance to execution
Connect visual geometry to the representation
that actually runs in the simulator.
Drag to compare Same object · Same camera · Same pose
- Vertices
- Surface faces
Visual and physical meshes may differ. Their association is recorded; the displayed edges come from the executed representation.
One asset. Two linked views.
Appearance and mesh, in motion
The textured surface and mesh share the same cloth displacement. This qualitative demonstration uses a new cloth solve on the archived mesh.
- 01
Construct
Connect appearance, geometry, and a compatible representation.
- 02
Calibrate
Observe contact and response. Revise supported controls.
- 03
Freeze
Fix geometry, representation, parameters, and numerical settings.
- 04
Test independently
Change the stimulus. Record a separate behavioral outcome.
Parameters are effective simulation controls. A separate runtime check is matched to the same configuration when forming the retained dataset.
Feedback in the loop
Observe. Revise. Test again.
A recorded filled-bag case: a no-contact diagnosis
leads to an effective density update from 25 to 50.
Feedback guides a supported configuration change. After calibration accepts the candidate, its configuration is frozen before independent testing.
ASSET 008563Evidence beyond calibration
Freeze the candidate. Change the test.
The same density-50 candidate.
Different stimuli reveal different behavior.
FROZEN CANDIDATE / 008563
Method replays are approximately 16.67× slower than physical time. Displayed metrics summarize the complete test.
Rejected after passing calibration
Independent testing adds evidence.
In the expanded bag study, 325 of 8,400 candidate evaluations pass calibration but fail the independent stimulus. Repeated methods and budgets share assets; the rejected evaluations involve 26 distinct bags.
Read the experiment in the manuscriptFull-pool evidence
Coverage after both checks
83,365 retained assets. All 485 categories.
View exact dataset counts
| Family | Categories | Generated | Physical | Runtime | Both |
|---|
Verification concerns the documented solver and interaction conditions. The 120-asset showcase is a curated view, not the full-pool distribution.
Grasping benchmark
A successful lift is only part of the story
Grasp, lift, and hold.
Evaluate task completion alongside integrity.
Tofu grasping · Simulation and real-world examples of damage and success. Qualitative comparison with video-guided reconstruction of key gripper actions.
Task completion and object integrity
RFW-S evaluates task completion and object integrity as separate outcomes. Intact success requires completing the grasp, lift, and hold task with a deformation-derived damage score below 0.25.
The damage score is a postprocessed deformation criterion; it does not measure physical cracks or coupled fracture dynamics.
The failure simulation uses a gripper trajectory reconstructed from selected video keyframes and an approximate damage model. Depth and gripper closure remain estimated; this qualitative comparison does not establish quantitative sim-to-real transfer accuracy.
Paper results / 340 held-out instances per method
Task completion and intact success
Aggregate benchmark outcomes, separate from the tofu demonstration. The split holds out instances and conditions, with base geometries shared across splits. The privileged reference uses damage information and five closure trials.
Read. Explore. Build on it.
Paper & resources
RobotFlyWheel-Soft
Everyday objects, with configuration and evidence.
Citation
Public release links and citation details will be added when available.