brukg

Robotics Simulators — Key Differentiators

Quick-reference summary of how five major robotics simulators compare on the dimensions that matter most for RL.

Aspect	Newton	Genesis	Gazebo	Isaac Sim/Lab	MuJoCo/MJX
Unique Strength	Multi-world tiled camera with 5 channels	Elastomer tactile + temperature + 9-axis IMU noise	Broadest environmental sensor suite	RTX tiled rendering at scale	Fastest batched physics + rich force sensors
Vision Quality	Good (raytraced)	Good (3 backends)	Moderate (Ogre2)	Excellent (RTX)	Basic (OpenGL)
Sensor Noise	None	Comprehensive	Per-sensor in SDF	Basic	None
Parallelism	Multi-world native	n_envs batched	Single world	Fully vectorized	MJX vmap

brukg

Robotics Simulator Sensor Stacks — Full Comparison Matrix

Comparison of sensor capabilities across Newton, Genesis, Gazebo, Isaac Sim / Isaac Lab, and MuJoCo / MJX. Compiled from source-code reading (Newton, Genesis) and official documentation (Isaac Lab, Gazebo, MuJoCo).

Sensor Type	Newton	Genesis	Gazebo	Isaac Sim / Isaac Lab	MuJoCo / MJX
RGB Camera	Warp raytraced, multi-world tiled	Rasterizer (OpenGL), Raytracer (LuisaRender), BatchRenderer (Madrona GPU)	OpenGL rasterizer, Ogre2	RTX ray-traced (Omniverse), Tiled camera	OpenGL offscreen (mj_render), MJX-Warp batch renderer
Depth Camera	Ray-hit distance per pixel	Raycaster-based depth camera	DepthCamera SDF sensor	RTX depth annotator, RayCasterCamera	mjr_readPixels depth buffer
Segmentation	Shape-index per pixel (uint32)	Not explicit (camera returns RGB only)	Segmentation camera plugin	Semantic/instance segmentation annotators	Not native (

brukg

Newton vs Genesis — CUDA Graph Showdown

Throughput results from a head-to-head benchmark on NVIDIA RTX 5090 Laptop GPU (24 GB). 500 timed steps + 50 warmup, dt=0.01s, PD position hold @ kp=1000.

When Newton's CUDA graph capture is enabled (--cuda-graph), it eliminates the kernel-launch CPU bottleneck and matches Genesis on throughput.

Scenario	Envs	Genesis FPS	Newton + CUDA Graph FPS	Winner
Franka	1	2.5K	5.2K	Newton 2.1×