The Impact of Time Step Frequency on the Realism of Robotic Manipulation Simulation for Objects of Different Scales

Minh Q. Ta, Holly Dinkel, Hameed Abdul-Rashid, Yangfei Dai, Jessica Myers, Tan Chen, Junyi Geng, Timothy BretlarXiv 2023 · DOI · arXiv

Abstract

Simulating the physics of small-scale robotic manipulation is notoriously difficult, often resulting in "jittery" or unrealistic behavior. This paper demonstrates that increasing the Time Step Frequency (TSF)—the number of times a physics engine updates per second—is critical for accurately simulating small objects. Through systematic experimentation in NVIDIA Isaac Sim, we identify stability thresholds for objects of varying scales and quantify the significant computational cost associated with higher-fidelity simulations.

Introduction

In robotics, the "Sim2Real" gap—the discrepancy between simulated and real-world behavior—is often driven by contact physics errors. In simulation, objects may unnaturally penetrate each other or "explode" out of a gripper due to numerical errors in the physics solver. These errors are highly sensitive to scale. While a physics engine might handle a large box perfectly at standard settings, a tiny electronic component might jitter uncontrollably.

Visual comparison of simulation accuracy across different scales and Time Step Frequencies (TSF).

This study investigates how the size of an object and the frequency of the simulation clock interact to affect the realism of robotic grasping and assembly.

Methodology

The researchers utilized NVIDIA Isaac Sim to model a work cell featuring two Universal Robots UR5e manipulators with Robotiq Hand-E grippers. The core methodology involved:

  • Varying Scale: Two electronic components, an "Inserted Part" (Volume: 178.17 $mm^3$) and a "Base Part" (Volume: 224.29 $mm^3$), were tested at three scales: original (1x), doubled (2x), and tripled (3x).
  • Varying TSF: A range of Time Step Frequencies (from 45 to 180 steps per second) was evaluated.
  • Ground Truth Comparison: A high-fidelity reference trajectory was established at 360 TSF. "Trajectory Error" for lower TSF runs was measured against this benchmark.
  • Solver Selection: The Temporal Gauss-Seidel (TGS) solver was used for local constraint resolution.

Key Results and Insights

The experiments revealed that small objects are "physics-expensive." The smaller the object, the higher the TSF required to maintain stability.

Stability thresholds for the Inserted Part across different scales and Time Step Frequencies.

Stability Thresholds

As shown in the results, the "Base Part" (the larger component) remained stable at a TSF as low as 45 across all scales. In contrast, the smaller "Inserted Part" at 1x scale required at least 145 TSF to avoid inaccurate behavior characterized by jittering and trajectory deviation. When scaled up to 3x, the same part became stable at a lower TSF of 120.

The Accuracy-Performance Trade-off

There is a clear trade-off between simulation fidelity and speed. Increasing TSF from 120 to 180 improved accuracy but increased the simulation runtime by nearly 80%, from approximately 10.5s to 18.5s. Furthermore, the relative scale between the manipulator and the object matters; halving the scale of the robot itself required an even higher TSF (180) to achieve accuracy for the original-sized part.

Conclusion and Future Work

This research highlights that TSF is a critical parameter for closing the Sim2Real gap in small-scale manipulation. However, the computational overhead of high TSF suggests a need for more efficient approaches. Future work should investigate adaptive time-stepping, where the simulation frequency increases only during complex contact events, and explore how TSF interacts with other physics parameters like friction and restitution to further enhance simulation realism.

Summary updated

BibTeX
@misc{ta2023impact,
  author = {Minh Q. Ta and Holly Dinkel and Hameed Abdul-Rashid and Yangfei Dai and Jessica Myers and Tan Chen and Junyi Geng and Timothy Bretl},
  title = {{The Impact of Time Step Frequency on the Realism of Robotic Manipulation Simulation for Objects of Different Scales}},
  year = {2023},
  eprint = {2310.08233},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.08233},
}
Figure