SIM1: Physics-Aligned Simulation for Robot Learning in Deformable Worlds

> "You can't surrender to something you don't understand." — Richard Feynman

This post is a detailed Chinese forum analysis of SIM1 (arXiv:2501.01778), a simulation framework for robotic manipulation of deformable objects like cloth, paper, and silk.

Why Deformable Objects Are Hard

Rigid objects (bricks, cups, tables) follow predictable Newtonian mechanics. Deformable objects do not: a crumpled sheet of paper retains creases, fabric has its own "personality" (silk slips, denim is stiff, wool bounces), and predicting how these materials respond to grasping and folding defies simple formulas. Most simulators are built for rigid bodies, so robots that master folding in simulation fail in reality — the classic reality gap. The paper's core thesis: *simulation fails not for being synthetic, but for being ungrounded.*

The SIM1 Pipeline

SIM1's core idea: don't build a simulation world from scratch — digitize the real world into one.

1. Scene Digitization

Using multi-view cameras and Neural Radiance Fields (NeRF), SIM1 reconstructs a metric-consistent digital twin — a 3D copy whose dimensions and geometry strictly match the real scene, not just its appearance.

2. Physics Calibration

Objects are modeled as elastic bodies (position-based elastic modeling, related to Position Based Dynamics) with parameters like Young's modulus, Poisson's ratio, and damping. Instead of hand-tuning, SIM1 automatically optimizes these parameters by minimizing the difference between simulated and real trajectories — essentially differentiable-simulation-style system identification for deformables.

3. Behavior Expansion

A diffusion-based trajectory generator produces new robot motion trajectories conditioned on initial and goal states, mimicking how diffusion models generate images. A learned quality filter discards physically implausible trajectories, leaving high-quality synthetic training data.

Key Results

• 1:15 equivalence ratio: Policies trained on pure synthetic data match real-data training using only 1/15 as many real demonstrations (e.g., ~1,000 real demos instead of ~15,000).
• 90% zero-shot transfer success: Skills learned entirely in the calibrated simulator succeed on first real-world execution, without domain randomization or domain adaptation.
• ~50% generalization gains: On real-world deployment, SIM1-trained policies generalize to new objects and conditions about 50% better than real-data-only baselines, attributed to the broader state-space coverage of generated trajectories.

Comparison with Related Approaches

• Domain randomization blindly randomizes simulator parameters hoping to cover reality; SIM1 *aligns* parameters via measurement instead of guessing.
• Classical system identification targets simple rigid systems; SIM1 handles complex deformable dynamics via learned calibration.
• Prior digital-twin work focuses on rigid or static scenes; SIM1 combines deformable-specific digital twins with large-scale data generation.

Limitations and Future Directions

• Requires a multi-view camera setup for scene digitization, complicating rapid deployment (e.g., home robotics).
• Physics calibration is an iterative optimization with unreported compute cost, potentially limiting fast adaptation.
• Diffusion-generated trajectories inherit biases from demonstrations, and the quality filter may miss subtly flawed trajectories.
• Future work: online learning after deployment, faster calibration (e.g., via meta-learning), and extension to liquids, granular media, and biological tissue.

References

1. Zhou, Y., Liu, H., & Jiang, X. (2025). SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds. arXiv:2501.01778. 2. Mildenhall, B., et al. (2020). NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV 2020. 3. Ho, J., Jain, A., & Abbeel, P. (2020). Denoising Diffusion Probabilistic Models. NeurIPS 2020. 4. Müller, M., et al. (2007). Position Based Dynamics. JVCIR, 18(2), 109–118. 5. Feynman, R. P. (1965). The Character of Physical Law. MIT Press.

*Based on arXiv:2501.01778, written in a Feynman-inspired style for zhichai.net.*