Summary

Proposes a function-preserving real-to-sim-to-real data generation framework that synthesizes demonstrations from reconstructed real-world assets with no teleoperated seed trajectory required. It specifically targets contact-rich tasks (insertion, assembly) where standard shape augmentation breaks geometric fit constraints.

Key Contributions

  • Identifies that standard shape-augmentation pipelines distort task-critical geometric interfaces, producing invalid contact relationships (fit mismatches, interpenetration)
  • Proposes constraint-guided asset augmentation that preserves functional/contact geometry while diversifying object shape
  • Generates synthetic demonstrations end-to-end from reconstructed assets with no teleoperated seed trajectory, targeting zero-shot real-to-sim-to-real transfer

Strengths

  • Directly targets a real, under-addressed failure mode of generic 3D data augmentation (geometric interface distortion) rather than generic visual domain randomization
  • Removing the teleoperated-trajectory requirement could meaningfully lower the cost of scaling synthetic data for precision/contact-rich tasks specifically

Weaknesses

  • “Function-preserving” constraints likely require task-specific geometric annotations or priors, limiting scalability to arbitrary new tasks without manual specification of what geometry must be preserved
  • With no teleoperated seed trajectories, action-label plausibility rests entirely on the trajectory optimizer/physics reasoning, a strong assumption for very high-precision insertion tasks

Open Questions

  • How well does policy performance trained purely on this synthetic data transfer to real contact-rich tasks compared to methods that still use a handful of real demonstrations?
  • Does the approach generalize beyond rigid-object insertion/assembly to deformable or articulated contact-rich tasks?

Significance

Adds a geometrically-principled alternative to the growing body of synthetic-data-for-VLA-training work, addressing a specific and previously under-scrutinized weakness (contact/interface distortion) in existing augmentation pipelines.