The direct answer
What does the paper report? EsaacSim is presented as an NVIDIA Isaac Sim add-on that can generate grayscale or Bayer RGGB event streams and synchronize RGB, APS, event, depth and IMU outputs through ROS 2. Its fixed v1 paper evaluates five resolutions at interpolation factors of 8×, 16× and 32× on one RTX 4060 8 GB GPU.
What do the headline numbers mean? The reported 6.98–27.28 ms grayscale and 7.58–29.16 ms Bayer ranges measure only the event-generation backend for one interval between rendered frames. The 240, 480 and 960 Hz figures are effective temporal sampling settings derived from a 30 Hz rendered cadence—not demonstrated Isaac Sim processing throughput.
What did RCI verify? RCI reviewed the fixed arXiv abstract and HTML v1, transcribed the reported configurations and boundaries, and ran zero simulations, GPUs, event cameras or robot workloads. The numbers remain author-reported.
The paper exposes a bounded, single-GPU backend evaluation across five resolutions and two output modes. It does not establish physical event-camera fidelity, end-to-end simulator throughput, multi-camera scaling, downstream robot perception performance or currently public source-code availability.
Five configurations, three interpolation settings
| Configuration | Resolution | Pixels | 8× effective sampling | 16× effective sampling | 32× effective sampling |
|---|---|---|---|---|---|
| DVS128 | 128 × 128 | 16,384 | 240 Hz | 480 Hz | 960 Hz |
| DAVIS240 | 240 × 180 | 43,200 | 240 Hz | 480 Hz | 960 Hz |
| ATIS | 304 × 240 | 72,960 | 240 Hz | 480 Hz | 960 Hz |
| DAVIS346 | 346 × 260 | 89,960 | 240 Hz | 480 Hz | 960 Hz |
| GenX320 | 320 × 320 | 102,400 | 240 Hz | 480 Hz | 960 Hz |
The effective sampling labels describe synthesized temporal resolution. They do not say that Isaac Sim rendered 240, 480 or 960 complete frames per second, and they do not represent physical event-camera output rates.
Allocated GPU memory is reported for each mode
| Sensor | 8× gray / Bayer | 16× gray / Bayer | 32× gray / Bayer |
|---|---|---|---|
| DVS128 | 51.4 / 51.4 MB | 55.4 / 55.4 MB | 64.0 / 64.1 MB |
| DAVIS240 | 135.5 / 135.6 MB | 146.0 / 146.2 MB | 167.8 / 168.0 MB |
| ATIS | 228.7 / 229.0 MB | 246.5 / 246.8 MB | 282.4 / 282.7 MB |
| DAVIS346 | 287.0 / 288.5 MB | 310.4 / 310.8 MB | 354.3 / 354.6 MB |
| GenX320 | 324.2 / 324.6 MB | 349.2 / 349.7 MB | 399.3 / 399.7 MB |
The reported allocated-memory range runs from 51.4 MB for DVS128 at 8× to 399.7 MB for GenX320 Bayer at 32×. The paper reports no more than 1.5 MB difference between Bayer and grayscale modes. RCI did not reproduce these measurements or audit GPU-allocation instrumentation.
6.98–29.16 ms covers the backend, not the whole stack
The paper reports grayscale event-stream generation between 6.98 and 27.28 ms and Bayer generation between 7.58 and 29.16 ms. It explicitly excludes scene loading, rendering initialization, ROS 2 communication, logging and file I/O from the benchmark.
That exclusion changes the allowable claim. The values can describe the paper's event-generation backend under its stated setup. They cannot be converted into end-to-end Isaac Sim frame rate, robot perception latency, real-time control capability or a cross-simulator ranking.
Four missing objects prevent a stronger claim
- No physical-sensor comparison: the fixed paper reports no quantitative comparison against recordings from a real event camera.
- No physics-grounded backend: the evaluated implementation is frame based; a physics-grounded backend remains future work.
- No downstream robot task: no localization, SLAM, detection or control workload is quantitatively evaluated.
- No confirmed public reproduction package: the audited arXiv surfaces expose neither a repository URL nor a raw result bundle. Mixed present/future release wording leaves current code availability unresolved.
Twelve disclosure checks
| Field | Status | Evidence | Boundary |
|---|---|---|---|
| fixed paper identity | disclosed | arXiv:2608.08522v1 and its 2026-08-09 submission date are fixed. | Later arXiv versions may change the text or results. |
| item license | CC-BY-SA-4.0 | The arXiv item declares Creative Commons Attribution-ShareAlike 4.0. | RCI's own dataset remains CC BY 4.0; linked paper content retains its stated license. |
| benchmark hardware | partially-disclosed | The evaluation names one NVIDIA GeForce RTX 4060 with 8 GB memory. | No multi-GPU, cross-driver or cross-host result matrix is reported. |
| benchmark scope | backend-only | The paper measures the event-generation backend. | Scene loading, initialization, rendering, ROS 2, logging and file I/O are excluded. |
| effective temporal sampling | disclosed-derived-setting | 30 Hz rendered cadence with 8×, 16× and 32× interpolation yields 240, 480 and 960 Hz temporal samples. | These values are not end-to-end Isaac Sim throughput or physical camera output rates. |
| real event camera comparison | absent | The paper reports no quantitative comparison against real event-camera recordings. | Simulation fidelity to a physical sensor is not established by this evaluation. |
| physics grounded backend | not-implemented | The evaluated backend is frame based; a physics-grounded backend is described as future work. | The paper does not demonstrate photoreceptor-level or circuit-level event-camera physics. |
| dynamic and multi camera workloads | not-evaluated | The reported study is limited to static or quasi-static scenes and one-camera configurations. | Dynamic scenes and multi-camera scaling require further evaluation. |
| downstream robot perception tasks | not-evaluated | No downstream detection, localization, SLAM or control task is quantitatively evaluated. | Backend timing does not establish robot-task accuracy or utility. |
| public source repository | unresolved-not-demonstrated | The v1 paper uses both release-oriented wording and future-tense public-release wording, but the audited arXiv surfaces expose no repository URL. | RCI does not claim that EsaacSim source code is publicly available as of this review. |
| public raw result bundle | not-found-in-audited-surfaces | The fixed paper surfaces do not expose a machine-readable raw timing/event-output bundle or analysis script. | The reported table values remain author-reported rather than independently recomputed. |
| rci independent execution | not-performed | RCI reviewed the fixed primary paper surfaces only. | RCI ran zero simulations, zero GPUs, zero event cameras and zero robot workloads for this brief. |
Method, rights and limits
RCI reviewed the fixed arXiv abstract and HTML v1 as primary sources, transcribed the paper's stated configurations, timing bounds, allocated-memory values and evaluation exclusions, and classified each statement by what the paper directly reports. RCI did not download or execute code, reproduce tables, inspect a physical event camera or generalize backend measurements to end-to-end simulator or robot performance.
- All performance and memory values are author-reported; RCI did not independently recompute them.
- The paper evaluates a frame-based backend on one RTX 4060 8 GB GPU.
- The evaluation excludes rendering and communication overhead, so the values are not full-system throughput.
- There is no quantitative physical event-camera comparison in the fixed v1 paper.
- No public repository or raw result bundle was confirmed from the audited arXiv surfaces.
- Later paper versions or an eventual code release may resolve some present boundaries.
The source paper declares CC BY-SA 4.0. This RCI evidence dataset and original editorial analysis are released under RCI's CC BY 4.0 policy; the linked paper and any future upstream code or assets retain their own licenses.
Primary sources: arXiv abstract and fixed HTML v1.
Download the evidence brief
Download the immutable JSON release and record-level CSV. Stable aliases are current JSON and current CSV.
Suggested citation: Robot Component Index. “EsaacSim Event-Camera Simulation: 5 Resolutions, 6.98–29.16 ms Backend Time, Zero Real-Sensor Comparisons.” RCI 040, version 0.1.0, 2026-08-11. https://robotcomponentindex.com/research/esaacsim-event-camera-simulation-evidence-brief/
For version-aware BibTeX and machine citation records, use Cite RCI. For evidence labels and correction rules, use the methodology.