Simulators

Simulators are the physics engines and rendering stacks where robot policies are trained, stress-tested, and iterated long before they touch hardware. They span rigid-body dynamics, contact-rich manipulation, photorealistic sensor simulation, and GPU-parallel rollouts that turn weeks of real-world data collection into hours of synthetic experience.

From an engineering standpoint, the simulator is a load-bearing dependency: its solver fidelity, contact model, and randomisation surface bound how well a policy can transfer to real hardware, and its throughput bounds how much you can afford to train. The wrong simulator quietly caps the ceiling of every downstream result — sim-to-real gaps, evaluation noise, and hidden modelling shortcuts all originate here.

When choosing between options, weigh physics fidelity (contact, friction, articulated bodies, deformables) against throughput (single-thread CPU vs. GPU-parallel) and sensor realism (ray-traced cameras, depth, tactile). Then check the surrounding ecosystem: ROS integration, asset libraries, RL wrappers, and active maintenance. Match the simulator to the task — dexterous contact work, photoreal perception, and large-scale RL each have different sweet spots.

Start here

If you are new to the space, MuJoCo is the most pragmatic entry point: fast, accurate contact dynamics, a small API surface, and the de facto baseline for robot-learning research.

Entries

• MuJoCo — Multi-joint dynamics with contact; fast, accurate physics widely used for RL research.

  simulatoropen sourceproduction ready

• NVIDIA Isaac Sim — GPU-accelerated robotics simulator on Omniverse with photorealistic rendering and PhysX.

  simulatorcommercialproduction ready

• Isaac Lab — Unified robot learning framework on Isaac Sim for RL, imitation learning, and motion planning.

  simulatorframeworkopen source

• Drake — Model-based design toolbox from TRI/MIT for planning, control, and rigorous dynamics analysis.

  simulatortoolopen sourceproduction ready

• Gazebo — Open-source robotics simulator with mature ROS integration and broad sensor support.

  simulatoropen sourceproduction ready

• PyBullet — Open-source physics engine (Bullet) with Python bindings, popular for prototyping and RL.

  simulatortoolopen source

• Habitat — Embodied AI platform optimised for high-throughput 3D navigation and instruction-following research.

  simulatoropen sourceresearch only

• SAPIEN — Physics-rich simulator with the PartNet-Mobility articulated object dataset.

  simulatortoolopen sourceresearch only

• Genesis — Universal differentiable simulator for robotics and embodied AI with cross-platform physics solvers.

  simulatoropen sourceresearch only

• Webots — Open-source robot simulator with mature educational and research workflows for mobile and manipulation robotics.

  simulatoropen sourceproduction ready

• CoppeliaSim — General-purpose robot simulation platform with rich scene scripting and broad manipulation benchmark usage.

  simulatortool

• CARLA — Open urban-driving simulator used for closed-loop autonomy and robustness testing at scale.

  simulatoropen sourcebenchmark

• AirSim — Photoreal simulation for drones and autonomous vehicles with configurable sensors and environments.

  simulatoropen source

• Brax — Differentiable, accelerator-native physics engine designed for high-throughput RL experimentation.

  simulatoropen sourceframework

• RaiSim — High-performance rigid-body simulator widely used for legged-locomotion research and control.

  simulatortoolcommercial