Simulator

Software that models in a computer how robots, objects, and sensors move and interact.

A simulator is software that recreates a robot and its environment inside a computer. It typically includes a physics engine, which advances the state forward each step by computing forces, collisions, and contact; a renderer, which generates camera images, depth, and other sensor data; scenes and assets, meaning the robot model and various objects; and an interface for algorithms to call, such as resetting the environment, executing one action, and returning observations and reward. In embodied AI, simulators are used to train reinforcement-learning policies, generate synthetic data, and run simulation-based evaluation, and their appeal is being cheap, safe, parallelizable, and reproducible. Common ones include MuJoCo (maintained by Google DeepMind, open-sourced in 2022), NVIDIA's Isaac Sim/Isaac Lab, PyBullet, Genesis, and SAPIEN. The core challenge is always the gap between it and reality.

ExampleThe ACT paper built two simulated bimanual tasks in MuJoCo, “transfer cube” and “bimanual insertion,” so others could reproduce the experiments without buying ALOHA hardware.

Also called

Simulation

Related

Physics Engine · Rendering · Sim-to-Real Transfer · MuJoCo (Multi-Joint dynamics with Contact) · NVIDIA Isaac Sim · Simulation-Based Evaluation

Sources

MuJoCo Documentation: Overview
NVIDIA Isaac Sim Documentation
Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT, arXiv 2304.13705)

As of

2026-09