GPU Robot Simulator Test: Genesis Hits 11.1 Million Physics Frames per Second
Robots are trained in GPU-powered simulators, but almost no one has checked how honest their physics is or whether their results can be reproduced. AIRI, MIPT and Shanghai AI Lab compared seven simulators on a single RTX 5070. Genesis, running a Franka robot arm, delivered about 11.1 million physics frames per second (FPS) across 131,072 environments. Only Genesis and MJX were fully repeatable, so FPS is not the only number to watch.
AI-processed from Habr AI; edited by Hamidun News
To stop a robot from dropping cups, it needs millions of attempts. On real hardware that is expensive and risky, so training is increasingly moved into a simulator, and the simulator runs on a graphics card. There, thousands of copies of the same scene with a robot run at once; each copy is called an environment.
Speed is measured in FPS. Rendering is turned off in these tests, so this is not frames shown on screen but the number of physics steps the simulator computes per second, summed across all environments. Genesis, running a Franka robot arm, delivered about 11.
1 million FPS across 131,072 environments. On average, that works out to roughly 85 physics steps per second for each copy of the scene. A team led by Dmitry Yudin from AIRI and MIPT, together with Shanghai AI Lab, built the GPUSimBench benchmark to answer two questions.
Does the physics stay plausible when hundreds of thousands of environments run on one card? And does the same experiment produce the same result when it is run again? The work was presented at the IROS 2026 conference.
How the test works
The comparison included seven GPU-accelerated simulators: Isaac Lab, ManiSkill, Genesis, Madrona, MuJoCo Warp, MJX and MuJoCo Playground. Platforms that parallelize only on the CPU, such as Gazebo and PyBullet, were not included. All measurements were made on one workstation with an Nvidia RTX 5070 graphics card and 12 GB of memory. Rendering was disabled, and the physics step was 0.01 s everywhere.
Scalability was tested on two scenes: a freely falling 3×3×3 stack of cubes, and a Franka Panda arm making random joint movements. To check realism, the authors built a real-world setup. A steel ball rolls down a wooden ramp at a 20° angle and hits a wall of 27 wooden cubes. After five seconds, the position of each cube is recorded. The scene is chaotic: even a tiny difference in initial conditions produces a different arrangement. Sixteen real runs defined the reference distribution, and sixteen parallel environments in the simulator produced its digital counterpart. The discrepancy between them was measured in centimeters.
Speed, realism and repeatability
On speed, the picture is as follows:
- Genesis leads for manipulators: about 11.1 million FPS on Franka with 131,072 environments.
- MJX and MuJoCo Warp are the record holders for environment count: 524,288 environments each with Franka.
- Madrona sustained 262,144 environments on falling cubes at about 481k FPS, but in this configuration it does not support manipulators.
- Isaac Lab delivered a balanced result: about 472k FPS on falling cubes and only 3.3 GB of GPU memory.
- MuJoCo Playground was the slowest: about 35 FPS on falling cubes at 256 environments, the maximum it supports. GPU acceleration alone does not guarantee speed.
On realism, the news is good: in the chaotic experiment, five of the seven simulators deviated from reality by roughly five centimeters. Basic contact physics works reasonably well. The exception is Madrona. Its XPBD solver in this configuration did not reproduce friction on the wooden surface, so the cubes barely slowed down and slid more than two meters. The gap between simulation and reality grew roughly 50-fold. Success-rate benchmarks would not catch this: a grasped cup still counts as grasped.
Repeatability was the most unexpected result. The random number generators were fixed, randomization was disabled, and the scenes were made identical. The team expected identical results but got four distinct modes:
- Isaac Lab and ManiSkill: environments within a single run differ by 4–5 cm, but the overall outcome of a run is reproducible.
- Genesis and MJX: fully deterministic in these tests, the most suitable mode for fair comparisons.
- Madrona and MuJoCo Warp: results diverge both between environments and between runs.
- MuJoCo Playground: environments within a run agree, but independent runs differ by about 1.4 cm.
What to do in practice, and where the limits are
Suppose you are comparing two control policies in Isaac Lab: you run one in the first half of the environments and the other in the second half. Since environments within a run differ by 4–5 cm, part of the difference may come from simulator noise rather than from policy quality. This is our own inference from the authors' table, not a direct claim by the authors.
If honest A/B comparisons and reproducibility matter to you, Genesis and MJX looked best in these tests. For the largest number of parallel environments with manipulators, MJX and MuJoCo Warp were the choice. For speed on a manipulator, Genesis was the choice.
As you increase the number of environments, watch not only FPS but also the stability of the results.
The limitations matter too. All tests ran on one consumer graphics card, and the real-world reference was built from a single scene with 16 trials. The conclusions about specific simulators apply to the configurations the authors used: for example, Madrona with XPBD might perform differently under other settings. The article and the benchmark code are published, so the results can be checked on your own hardware.
- Nov 21, 2026Nvidia doubled profit, but shares fell 2% over the search for a new CEO
- Nov 21, 2026Nous Research fine-tuned China's Qwen3.8, and propaganda in its answers dropped from 22% to 1%
- Nov 21, 2026Samsung Wants to Buy Modine, a Data Center Cooling Supplier, for Nearly $6 Billion
- Nov 21, 2026Reuters: China's $3.2B AI Data Center Buildout Relies on Nvidia Chips as Huang Steps Down
Key Questions Answered
How the test works?
The comparison included seven GPU-accelerated simulators: Isaac Lab, ManiSkill, Genesis, Madrona, MuJoCo Warp, MJX and MuJoCo Playground. Platforms that parallelize only on the CPU, such as Gazebo and PyBullet, were not included. All measurements were made on one workstation with an Nvidia RTX 5070 graphics card and 12 GB of memory. Rendering was disabled, and the physics step was 0.01 s everywhere.
What are the key details on Speed, realism and repeatability?
On speed, the picture is as follows: Genesis leads for manipulators: about 11.1 million FPS on Franka with 131,072 environments. MJX and MuJoCo Warp are the record holders for environment count: 524,288 environments each with Franka. Madrona sustained 262,144 environments on falling cubes at about 481k FPS, but in this configuration it does not support manipulators. Isaac Lab delivered a balanced result: about 472k FPS on falling cubes and only 3.3 GB of GPU memory.
What to do in practice, and where the limits are?
Suppose you are comparing two control policies in Isaac Lab: you run one in the first half of the environments and the other in the second half. Since environments within a run differ by 4–5 cm, part of the difference may come from simulator noise rather than from policy quality. This is our own inference from the authors' table, not a direct claim by the authors. If honest A/B comparisons and reproducibility matter to you, Genesis and MJX looked best in these tests.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.