# Simulation (in AI and robotics)

> Source: https://aiwiki.ai/wiki/simulation
> Updated: 2026-09-25
> Fact-checked: 2026-09-25
> Categories: Embodied AI, Physical AI, Reinforcement Learning, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Simulation (in AI and robotics)." aiwiki.ai, 25 Sept 2026. https://aiwiki.ai/wiki/simulation
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Simulation** in [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) and [robotics](https://aiwiki.ai/wiki/robotics) is the use of computational physics, rendering, and procedural environments to recreate a synthetic version of a physical or virtual world inside which [AI agents](https://aiwiki.ai/wiki/ai_agents) can perceive, act, learn, and be evaluated. It is one of the main ways modern robots and [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) agents are trained: virtual robots gather large amounts of experience inside simulators, and the resulting policies are then deployed on real hardware. The scale advantage is large. ETH Zurich and NVIDIA researchers reported in 2021 that thousands of simulated quadrupeds running in parallel on a single workstation GPU could learn a flat-terrain walking policy in under four minutes, [45] and the [Genesis](https://aiwiki.ai/wiki/genesis_simulator) project reported 43 million frames per second for a Franka arm scene on a single RTX 4090, which it described as 430,000 times faster than real time. [9] Most contemporary work on [humanoid robots](https://aiwiki.ai/wiki/humanoid_robot), autonomous vehicles, dexterous manipulation, and embodied agents leans on simulation for training, testing, or both.

The term means something specific here. In statistics and physics, "simulation" can mean Monte Carlo sampling or finite-element solvers run for engineering. In the AI and robotics context the focus is narrower: physics engines and 3D environments wired into [machine learning](https://aiwiki.ai/wiki/machine_learning) pipelines, with goals like data collection, [domain randomization](https://aiwiki.ai/wiki/domain_randomization), [sim-to-real](https://aiwiki.ai/wiki/sim_to_real_transfer) transfer, and the training of generalist policies. Since 2025 much of the new GPU-parallel robotics work has been developed in the open, including [MuJoCo](https://aiwiki.ai/wiki/mujoco) Warp (Google DeepMind and NVIDIA) and [NVIDIA Newton](https://aiwiki.ai/wiki/nvidia_newton) (NVIDIA, Google DeepMind and Disney Research), which uses MuJoCo Warp as a key solver. [29][31] The frontier has also expanded to include generative models that learn the simulator itself from video, blurring the line between traditional rigid-body engines and neural [world models](https://aiwiki.ai/wiki/world_model).

## Why does simulation matter for AI and robotics?

Real robots are expensive, slow, and break. A single [Boston Dynamics](https://aiwiki.ai/wiki/boston_dynamics) Spot or a [Unitree H1](https://aiwiki.ai/wiki/unitree_h1) humanoid costs tens of thousands of dollars; replacement parts can take weeks; and a fall during a learning episode might end an experiment for a day. Simulation sidesteps many of those constraints.

A simulator gives researchers cheap, parallelizable, safe, and resettable data. You can run thousands of robots on a single GPU, step them much faster than real time, and reset to any prior state on demand. Bad policies break virtual robots without consequences. Curricula are easy to design because you can vary anything: gravity, friction, lighting, the mass of an object, the geometry of a kitchen. NVIDIA's own summary of the case lists starting development before hardware is ready, running more experiments in less time, identifying failures before they damage equipment, and evaluating behavior across repeatable scenarios and edge cases. [29]

There are several distinct reasons the field leans on simulation:

- **Cheap data collection.** Physical data collection is slow and costly, while a GPU-parallel simulator like [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab) or Brax steps thousands of environments at once. The Brax authors reported millions of simulation steps per second on the MuJoCo Ant task from a single accelerator. [7]
- **Safe exploration.** RL agents need to try bad actions to learn. A walking robot has to fall over many times before it stops falling. Letting that happen in simulation avoids hardware damage and human risk.
- **Counterfactuals.** Simulators let you ask what-if questions. What if the door were heavier, the floor more slippery, the camera placed 5 cm to the left? In the physical world, comparing those conditions cleanly is difficult; in simulation it is a parameter sweep.
- **Reproducibility.** Real-world experiments are hard to reproduce because lighting, calibration, and wear vary. A simulated scenario can be reset and rerun with the same settings, and engines such as MuJoCo are known for a deterministic pipeline, so different research groups can compare methods under matched conditions. [37]
- **Coverage.** Edge cases that are rare in reality (a child running into the road, a robot dropping a glass on a tile floor) can be sampled at will. Waymo, for example, describes its driver "navigating billions of miles in virtual worlds, mastering complex scenarios long before it encounters them on public roads." [50]
- **Speed.** Massively parallel GPU simulation changed which algorithms are practical. NVIDIA's Isaac Gym report measured 2-3 orders of magnitude faster training than conventional setups that pair a CPU simulator with a GPU for the neural network, [5] and the Brax paper claimed a 100-1000x improvement in the speed and cost of RL training by putting physics and the optimizer on the same accelerator. [7]

All of this comes with the central, unsolved tradeoff: a simulator is a model, and models are wrong. Bridging the gap to the real world is the central engineering problem of the field, addressed mainly through domain randomization, system identification, and [domain adaptation](https://aiwiki.ai/wiki/domain_adaptation).

## What are the major physics engines for AI and robotics?

The physics engine is the core of any simulator. It computes how bodies move, collide, deform, and interact with actuators and sensors. The engines listed below are the ones most commonly cited in robotics and RL research; each has different tradeoffs in accuracy, speed, parallelism, and ergonomics.

| Engine | Origin | License | Strengths | Common use |
| --- | --- | --- | --- | --- |
| [MuJoCo](https://aiwiki.ai/wiki/mujoco) | Todorov, Erez and Tassa (IROS 2012); DeepMind acquired it in October 2021 and completed open-sourcing in May 2022 | Apache 2.0 | Fast, accurate contact-rich rigid-body dynamics; automatic differentiation through MJX; GPU batching through MJX and MuJoCo Warp | Continuous-control RL benchmarks; humanoid locomotion; manipulation |
| Bullet / [PyBullet](https://aiwiki.ai/wiki/pybullet) | Erwin Coumans (Sony, AMD, Google, NVIDIA) | zlib | Mature collision detection; Python-friendly; URDF support | Hobbyist robotics, classic OpenAI Gym tasks |
| [Gazebo](https://aiwiki.ai/wiki/gazebo_simulator) | 2002 (Player Project), then Willow Garage and Open Robotics | Apache 2.0 | Tight ROS integration; sensor models; large robotics community | ROS-based simulation, system integration testing |
| Isaac Sim and [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab) | NVIDIA, built on [Omniverse](https://aiwiki.ai/wiki/nvidia_omniverse) and PhysX; Isaac Lab 3.0 adds Newton and kit-less backends | Isaac Sim source Apache 2.0 (Omniverse Kit and assets under separate NVIDIA terms); Isaac Lab BSD-3-Clause, some parts Apache 2.0 | Photoreal rendering; GPU parallel; OpenUSD scene format | Industrial robotics, humanoids, large-scale RL |
| [Newton](https://aiwiki.ai/wiki/nvidia_newton) | NVIDIA, Google DeepMind and Disney Research; announced March 2025; Linux Foundation project since September 2025 | Apache 2.0 | Multi-solver GPU engine built on NVIDIA Warp and OpenUSD; MuJoCo Warp, Kamino, VBD and implicit MPM solvers | Isaac Lab physics backend, contact-rich manipulation, deformables |
| Brax | Google, 2021 | Apache 2.0 | Differentiable; written in JAX; massive parallelism on TPU/GPU | Differentiable RL, learned dynamics, JAX pipelines |
| Drake | Started by Russ Tedrake's Robot Locomotion Group at MIT CSAIL; core development now led by TRI | BSD-3-Clause | Rigorous multibody dynamics; strong contact mechanics; optimization tooling | High-fidelity research, control theory, manipulation |
| Genesis (now Genesis World) | Academic consortium, December 2024; development now supported by Genesis AI | Apache 2.0 | Multi-physics (rigid, FEM, MPM, SPH, PBD); Python; very fast | Generative robotics workflows, embodied AI research |
| Webots | Cyberbotics (originally EPFL, 1996); open-sourced December 2018 | Apache 2.0 | Educational use, large robot library, scripted scenarios | Teaching, RoboCup, prototyping |

### MuJoCo

[MuJoCo](https://aiwiki.ai/wiki/mujoco) (Multi-Joint dynamics with Contact) was published by Emanuel Todorov, Tom Erez and Yuval Tassa at IROS 2012 and became a standard physics engine for academic continuous-control RL. [1] Most of the canonical benchmark tasks (HalfCheetah, Humanoid, Ant, the OpenAI Gym MuJoCo suite) use it. DeepMind announced in October 2021 that it had acquired MuJoCo and was making it freely available as a precompiled library, [61] and in May 2022 it reported that open-sourcing was complete, with the full codebase on GitHub; [2] the repository is published under the Apache 2.0 license. The May 2022 post recalled that DeepMind had "committed to developing and maintaining MuJoCo as a free, open-source, community-driven project with best-in-class capabilities." [2]

The modern MuJoCo ecosystem includes MJX, a JAX API whose JAX re-implementation runs on NVIDIA and AMD GPUs, Apple Silicon, and Google Cloud TPUs and is automatically differentiable through JAX, [34] and MuJoCo Warp (MJWarp), a GPU implementation written in [NVIDIA Warp](https://aiwiki.ai/wiki/nvidia_warp) and developed jointly by NVIDIA and Google DeepMind. [33] At its March 2025 announcement NVIDIA said MuJoCo-Warp gave "more than a 70x acceleration for humanoid simulations and a 100x speedup for in-hand manipulation tasks"; the post did not name the baseline. [26] MuJoCo 3.5.0 (February 12, 2026) marked MuJoCo Warp's official release, [35] and MuJoCo and MJWarp now share version numbers, with 3.14.0 released on September 22, 2026. [35][36] MuJoCo also underpins [RoboCasa](https://aiwiki.ai/wiki/robocasa) (through the robosuite framework) and MuJoCo Playground, whose environments were built on MJX and now also train with MuJoCo Warp. [14][60]

### PyBullet and Bullet

Bullet is a zlib-licensed physics SDK written by Erwin Coumans, who worked for Sony Computer Entertainment US R&D from 2003 to 2010, then AMD, then Google until 2022, and now works for NVIDIA. [3][58] PyBullet is the Python binding that turned it into a standard tool for RL research; the Bullet repository itself recommends PyBullet for robotics and reinforcement learning work. [3] It is mature, well-documented, and a comfortable starting point for anyone new to robotics simulation, though for GPU parallelism it has been overshadowed by MuJoCo MJX and MuJoCo Warp, Brax, Isaac Lab, and Genesis. NVIDIA authors writing in 2026 describe PyBullet as "a useful CPU baseline for quick prototyping." [37]

### Gazebo

Gazebo started in 2002 as a component of the Player Project, became an independent project supported by Willow Garage in 2011, and has been stewarded by the Open Source Robotics Foundation (now Open Robotics) since 2012. [4] Its tight integration with ROS made it the default simulator for many industrial and academic robotics groups for over a decade. After more than 15 years of development the project was rebuilt as a collection of loosely coupled libraries under the name Ignition; in 2022 a trademark obstacle over the name "Ignition" led Open Robotics to rebrand the new libraries as Gazebo, with the original monolithic simulator now called Gazebo Classic. [59] Gazebo Classic reached end of life in January 2025. [59]

### NVIDIA Isaac

NVIDIA Isaac is the umbrella name for NVIDIA's robotics stack. It includes [Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim), a photorealistic GPU simulator built on Omniverse with PhysX physics and RTX rendering; [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab), the RL training framework that succeeded the deprecated Isaac Gym and whose documentation provides migration guides from IsaacGymEnvs, OmniIsaacGymEnvs, and Orbit; [5][6][46] and [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t), NVIDIA's humanoid robot foundation model effort. Isaac Sim uses OpenUSD as its core scene and data layer and simulates cameras, depth, lidar, and radar sensors. [37] Its source code is published on GitHub under Apache 2.0, with the Omniverse Kit SDK and 3D assets it depends on covered by separate NVIDIA terms; [41] the GitHub release history begins with Isaac Sim 5.0.0 (August 8, 2025) and reached 6.1.0 on September 10, 2026. [41]

Isaac Lab 3.0, released as Early Access on September 16, 2026, decouples Isaac Lab from Isaac Sim: the same task can run on Isaac Sim PhysX, a kit-less OVPhysX backend, or Newton with MuJoCo Warp, with rendering and visualization chosen separately. NVIDIA targets general availability toward the end of October 2026. [40]

### Brax

Brax is a differentiable rigid-body physics engine written in JAX, released by Google researchers (Freeman, Frey, Raichuk, Girgin, Mordatch, Bachem) in 2021 and published in the NeurIPS 2021 Datasets and Benchmarks track. [7] Because the simulator is JAX-traceable, environment dynamics, neural networks, and the optimizer all compile together and run on the same accelerator. The authors reported that Brax trains locomotion and dexterous manipulation policies "in seconds to minutes using just one modern accelerator." [7] Brax is also a natural home for differentiable physics research, where gradients flow through the dynamics into policies.

### Drake

The Drake project was started by Russ Tedrake and members of the Robot Locomotion Group at MIT CSAIL, and its core development is now led by the Toyota Research Institute. [8] It is more conservative than the GPU-first simulators above. Drake invests heavily in numerically robust contact mechanics, hydroelastic contact models, and a systems framework that integrates well with optimization-based control. It is C++ with Python bindings and is used in research where fidelity matters more than raw simulation throughput, such as control of humanoids, dexterous manipulation, and academic underactuated robotics. NVIDIA authors call it "the gold standard if you need contact-implicit trajectory optimisation and rigorous numerics rather than throughput," [37] and Newton's hydroelastic contact model is explicitly inspired by Drake's; TRI is also partnering with the Newton project on solver development and contact modeling. [32]

### Genesis

Genesis was announced on December 18, 2024 by Zhou Xian of Carnegie Mellon University, the first-listed core contributor, who described it as the result of "a 24-month large-scale research collaboration involving over 20 research labs." [44][9] It is an Apache 2.0 Python simulator whose physics engine integrates rigid-body, MPM, SPH, FEM, PBD, and stable-fluid solvers, paired with a generative data engine that turns natural-language descriptions into data; at launch only the physics engine and simulation platform were open-sourced, with the generative framework to be rolled out later. [9] The project reported that a Franka manipulation scene ran at 43 million frames per second on a single RTX 4090, roughly 430,000 times faster than real time, [9] and the launch announcement claimed the engine was "10-80x faster than existing GPU-accelerated stacks like Isaac Gym and MJX." [44] These are the developers' own benchmarks; independent comparisons across diverse workloads are the better guide.

The academic project has since become a company platform. Its README now calls it Genesis World and says development "is now officially supported by Genesis AI." [42] [Genesis AI](https://aiwiki.ai/wiki/genesis_ai) released Genesis World 1.0 on May 27, 2026, bundling the physics platform with Nyx, a path-traced renderer built for robotics, and Quadrants, a Python-to-GPU compiler forked from Taichi. [42][43] The company says it now uses simulation mainly as an evaluation engine for its robot foundation models and that its simulation-based evaluation "correlates strongly with on-hardware rollouts." [43]

### Webots

Webots was designed at EPFL in 1996 as a research tool for mobile robotics, commercialized by Cyberbotics from 1998 onward, and open-sourced under Apache 2.0 in December 2018. [10] It has a strong educational and competition footprint (RoboCup, university courses) and a polished GUI, with bindings for C, C++, Python, Java, MATLAB, and ROS.

## What simulators are used for embodied AI?

A second class of simulators sits on top of physics engines and provides large 3D environments populated with rooms, objects, and tasks. These are designed for embodied AI: agents that navigate and manipulate inside human environments. Photorealism, scene diversity, and task variety matter more here than raw physics throughput.

| Simulator | Lead organization | First released | Underlying engine | Focus |
| --- | --- | --- | --- | --- |
| Habitat | Meta AI (FAIR) | 2019 (Habitat 1.0) | Custom; Bullet for physics | Indoor navigation, embodied agents |
| Habitat 3.0 | Meta AI | October 2023 | Same lineage | Human-robot collaboration, social rearrangement |
| AI2-THOR | Allen Institute for AI | 2017 | Unity | Household interaction, navigation, manipulation |
| ManipulaTHOR | Allen Institute for AI | 2021 | Unity (AI2-THOR) | Mobile manipulation with a robotic arm in indoor scenes |
| iGibson and OmniGibson | Stanford | 2020 (iGibson 1.0) | Bullet (iGibson), then NVIDIA Omniverse (OmniGibson) | Interactive household tasks, BEHAVIOR benchmarks |
| BEHAVIOR-1K | Stanford | 2022 | OmniGibson | 1,000 everyday household activities |
| [RoboCasa](https://aiwiki.ai/wiki/robocasa) | UT Austin and NVIDIA (Nasiriany et al., RSS 2024) | June 2024 | MuJoCo via robosuite | Kitchen tasks for generalist robot policies |
| ManiSkill / ManiSkill3 | UC San Diego (Su Lab) | 2021; ManiSkill3 in 2024 | SAPIEN | GPU-parallel manipulation benchmark |
| ProcTHOR | Allen Institute for AI | 2022 | Unity | Procedurally generated 10K houses |

Habitat from Meta AI emphasizes fast navigation in photorealistic indoor scans (initially Matterport3D, Replica, and Gibson). [11] Habitat 3.0, released in October 2023, added humanoid avatars and human-in-the-loop infrastructure that lets real people interact with simulated robots via mouse and keyboard or a VR interface, opening up tasks like social rearrangement and social navigation. [12]

AI2-THOR from the Allen Institute for AI (AI2) takes the opposite approach: hand-modeled rooms in Unity with carefully crafted interactions. [13] Pour water into a kettle, place it on a stove, watch it boil. ManipulaTHOR added a robotic arm; ProcTHOR procedurally generated 10,000 houses. AI2's MolmoSpaces effort unifies over 230,000 indoor scenes and more than 130,000 object models with over 42 million annotated robotic grasps, and moves away from Unity's simplified "magic grasps" to physics engines such as MuJoCo. [55]

RoboCasa was published at Robotics: Science and Systems in 2024 by Soroush Nasiriany, Ajay Mandlekar, Yuke Zhu, and collaborators. [14] It is built on MuJoCo via the robosuite framework and focuses on kitchen environments with thousands of 3D assets, some produced with text-to-3D and text-to-image models. [14] The follow-up RoboCasa365 (v1.0, February 18, 2026) covers 365 everyday tasks across 2,500 kitchen environments, with more than 600 hours of human demonstrations and more than 1,600 hours of synthetically generated demonstrations. [54] ManiSkill3, from the Su Lab at UCSD, runs on SAPIEN and reports up to 30,000+ FPS with rendering in benchmarked environments. [15]

## What simulators are used for autonomous driving?

Self-driving research has its own simulator ecosystem because the relevant physics (large vehicles, road surfaces, traffic) and the relevant tasks (perception in adversarial conditions, multi-agent prediction) are quite different from indoor robotics.

| Simulator | Origin | Engine | Notes |
| --- | --- | --- | --- |
| CARLA | Intel Labs, Toyota Research Institute and Computer Vision Center (Barcelona), 2017 paper | Unreal Engine | Open-source, leading academic AV simulator |
| AirSim | Microsoft Research, 2017 | Unreal Engine, Unity plugin | Drones and ground vehicles; Microsoft said in July 2022 it would archive AirSim and shift to Project AirSim |
| NVIDIA AV simulation | NVIDIA, on Omniverse | Omniverse NuRec neural reconstruction, Cosmos world models | Closed-loop testing through the open AlpaSim framework |
| LGSVL (SVL) Simulator | LG | Unity | Active development suspended January 1, 2022 |
| SimulationCity and Waymo World Model | Waymo, internal | Proprietary; the 2026 world model is built on Genie 3 | Synthesizes full journeys; generates rare events for testing |

CARLA was introduced in the paper *CARLA: An Open Urban Driving Simulator* by Dosovitskiy, Ros, Codevilla, Lopez, and Koltun (Intel Labs, Toyota Research Institute, and the Computer Vision Center in Barcelona) at the first Conference on Robot Learning (CoRL) in 2017. [16] It is the standard academic simulator for autonomous urban driving, with a flexible sensor suite and configurable environmental conditions. AirSim was created at Microsoft Research in 2017; in July 2022 Microsoft announced that it would stop updating the original AirSim, archive it, and focus on a commercial product, Microsoft Project AirSim, for the aerospace industry. [17][51] As of September 2026 the GitHub repository is still online and not marked as archived. [17] LG suspended active development of its SVL Simulator as of January 1, 2022. [52]

NVIDIA's current AV simulation offering centers on Omniverse NuRec, which reconstructs interactive 3D scenes from real driving data, Cosmos world models, and AlpaSim, an open framework for closed-loop testing over "millions of virtual miles." [53] Waymo's internal SimulationCity, introduced in July 2021, automatically synthesizes entire journeys to assess the Waymo Driver, drawing on the more than 20 million autonomous miles the fleet had then driven. [49] In February 2026 Waymo introduced the Waymo World Model, a generative model built on Google DeepMind's [Genie 3](https://aiwiki.ai/wiki/genie_3) and adapted for driving, which can simulate rare events such as a tornado or an encounter with an elephant. [50]

## How does GPU-accelerated parallelism speed up simulation?

The most important shift in the last few years has been the move from single-threaded CPU simulators to massively parallel GPU simulators. The pattern is the same in each project: instead of stepping one environment at a time, batch thousands of environments together and step them all on the GPU, keeping physics and policy computation on the same device to avoid data-transfer overhead. [29] NVIDIA itself cautions that "the benefit depends on the workload, batch size, contact complexity, and hardware, so developers should benchmark their own setup." [29]

- **Isaac Gym and Isaac Lab.** Isaac Gym, released as a preview by NVIDIA, popularized GPU-resident simulation for RL; both physics and policy training reside on the GPU and exchange data through PyTorch tensors. [5] NVIDIA now lists Isaac Gym as deprecated legacy software and points developers to [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab), [46] which succeeded the older IsaacGymEnvs, OmniIsaacGymEnvs, and Orbit codebases. [6]
- **MJX.** A JAX API for MuJoCo that runs on any hardware the XLA compiler supports. It is widely used in research because it preserves MuJoCo semantics while running on accelerators, and its dynamics are automatically differentiable through JAX. [34]
- **Brax.** A pure-JAX rigid-body engine designed from day one for accelerators. The original paper claimed a 100-1000x improvement in the speed and cost of RL training and policies trained in seconds to minutes on one accelerator. [7]
- **Genesis.** Reports throughput in the millions of FPS on commodity GPUs with multi-physics support. [9]
- **MuJoCo Warp.** Introduced in March 2025 alongside the Newton announcement and officially released in MuJoCo 3.5.0 in February 2026. [26][35] NVIDIA's benchmark figures against MJX have grown with each release: up to 152x faster for locomotion and 313x for manipulation on a GeForce RTX 4090 at the September 2025 Newton Beta, [31] and 252x and 475x on the RTX PRO 6000 Blackwell Series at the March 2026 Newton 1.0 release. [32] These are vendor benchmarks.
- **Newton.** Uses MuJoCo Warp as its main rigid-body solver; NVIDIA reported that the Newton Beta backend gave up to 65% faster in-hand dexterous manipulation in Isaac Lab than PhysX. [31]
- **ManiSkill3.** Built on SAPIEN, with up to 30,000+ FPS in benchmarked environments that include rendering; the authors report 10-1000x faster simulation with rendering than other platforms. [15]

What changed in practice is that an algorithm like PPO that needed many CPU days now finishes in minutes to hours. The 2021 paper *Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning* by Rudin, Hoeller, Reist, and Hutter trained ANYmal flat-terrain policies in under four minutes and rough-terrain policies in twenty minutes on a single workstation GPU, then transferred them to the real robot. [45] The same recipe (thousands of parallel environments trained with the rsl_rl library) now ships as a stock Isaac Lab example for the ANYmal-D robot. [31]

## How does the modern GPU robotics simulation stack fit together?

By 2026 NVIDIA was describing its GPU robotics stack as a set of distinct layers. NVIDIA's "Physics Simulation for Robotics" use-case page, promoted by the NVIDIA Robotics account on September 24, 2026, describes a stack that "separates task definition, physics computation, GPU acceleration, and policy training so developers can select the components their workflow requires." [29][30] The pieces, and how they depend on one another, are:

| Layer | Component | What it does | Maintainers |
| --- | --- | --- | --- |
| Model format and CPU physics | [MuJoCo](https://aiwiki.ai/wiki/mujoco) | The core C physics library and the MJCF model format; simulates robot dynamics but does not include policy-training workflows [29] | Google DeepMind |
| JAX batching | MJX | JAX API over MuJoCo implementations; the JAX version runs on NVIDIA and AMD GPUs, Apple Silicon, and TPUs and is differentiable through JAX [34] | Google DeepMind |
| GPU physics kernels | MuJoCo Warp (MJWarp) | Runs compatible MuJoCo physics in NVIDIA Warp for batched GPU simulation; can be used directly, through MJX, through Newton in Isaac Lab, or through mjlab [29][36] | Google DeepMind and NVIDIA, as part of the Newton project [36] |
| Kernel framework | [NVIDIA Warp](https://aiwiki.ai/wiki/nvidia_warp) | Python framework for writing GPU simulation and spatial-computing kernels and connecting them to machine-learning frameworks [29] | NVIDIA |
| Physics engine | [Newton](https://aiwiki.ai/wiki/nvidia_newton) | Open-source engine built on Warp and OpenUSD that separates model, state, controls, contacts, and solver; MJWarp is a key rigid-body solver [29] | NVIDIA, Google DeepMind, Disney Research; Linux Foundation project [38] |
| Authoring and testing | [Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim) | Authoring environment for robot assets that can run on multiple physics backends, and software-in-the-loop testing of robot software with Newton as the physics backend [29] | NVIDIA |
| Robot learning | [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab) | Environment configuration, policy training, evaluation, and transfer; can use Newton as a physics backend [29] | NVIDIA |

Newton is a framework rather than a single solver. Its solvers include SolverMuJoCo and SolverFeatherstone for articulated rigid bodies in generalized coordinates; SolverSemiImplicit, SolverXPBD, and Disney Research's Kamino in maximal coordinates; SolverVBD, an implicit solver for rigid bodies, particles, cloth, and soft bodies; SolverImplicitMPM for particle-based continuum materials; and SolverStyle3D for cloth. Because the solvers differ in their support for articulations, deformables, contacts, and differentiation, "the appropriate solver depends on the physical system being modeled." [37] Inside Isaac Lab, the task definition, PPO loop, observations, and rewards stay identical when the physics backend changes, so developers can author an environment once and validate it on different engines. [32] Newton 1.0 was announced as generally available at GTC on March 16, 2026, [32] and the project shipped releases through v1.6.0 on September 10, 2026. [39]

Differentiability is uneven across this stack. MJX is differentiable through JAX, but MJWarp's documentation says differentiability via Warp "is not yet available." [34][36] Newton is described as differentiable at the engine level, with capabilities that vary by solver. [37]

### When does GPU simulation help?

GPU simulation trades latency for throughput. MuJoCo's documentation says MJWarp is optimized for "the total number of simulation steps per unit time" whereas CPU MuJoCo is optimized for the time of a single step, and that "a simulation step with MJWarp will be less performant than a step with MuJoCo for the same simulation." [33] MJWarp therefore suits workloads that need very large numbers of samples, such as reinforcement learning, while CPU MuJoCo remains more useful for real-time uses such as model predictive control or simulation-based teleoperation. [33] MJWarp scales better than MJX for scenes with many geoms or degrees of freedom, but not as well as CPU MuJoCo for single large kinematic trees; the MuJoCo team lists single connected mechanisms beyond about 60 degrees of freedom as an active priority. [33] CPU users can also batch work through `mujoco.rollout`, although frequent host-to-device transfers can then become the bottleneck. [33]

NVIDIA's own guidance is similar: moving a MuJoCo workflow to the GPU "can be valuable when training or evaluation requires many parallel environments and simulation throughput has become a bottleneck," and teams should "validate both performance and expected behavior before fully migrating the workflow." [29]

### Rigid-body versus multiphysics simulation

NVIDIA recommends rigid-body simulation "when the task primarily involves articulated robots and solid objects," and coupled multiphysics "when success depends on deformable materials, cables, fluids, tactile interactions, or other complex physical behavior." [29] Newton handles the second case by running specialized solvers side by side. In a standalone Newton demo, a Featherstone solver drives a Franka arm while a VBD solver simulates cloth, with one-way coupling from robot to cloth; NVIDIA reports about 30 FPS on an RTX 4090 with penetration-free contact and more than 300x higher performance than GPU-IPC. [31] By the 1.0 release, Newton's VBD solver covered cables, cloth, and volumetric rubber parts, the implicit MPM solver handled granular material for rough-terrain locomotion, and both could be coupled explicitly with MuJoCo Warp. [32] For tight-tolerance contact, Newton 1.0 added signed-distance-field collision built from CAD meshes and hydroelastic contacts that model a pressure distribution over a contact patch instead of a set of points. [32]

### Case studies

NVIDIA's use-case page highlights three workflows. [29] The underlying technical posts give the details.

| Case | Organizations | Physics setup | Result reported |
| --- | --- | --- | --- |
| Quadruped locomotion | ETH Zurich Robotic Systems Lab, NVIDIA | ANYmal-D policy trained in Isaac Lab with the Newton backend; Sim2Sim check in PhysX-based Isaac Lab | Policy deployed directly to a physical ANYmal, which walked [31] |
| GPU rack assembly | [Skild AI](https://aiwiki.ai/wiki/skild_ai) | Isaac Lab with Newton; SDF collision and hydroelastic contact replace MuJoCo Warp's native contact pipeline | Training RL policies for connector insertion, board placement, and fastening [32] |
| Refrigerator water-hose insertion | [Samsung](https://aiwiki.ai/wiki/samsung), Lightwheel | MuJoCo Warp two-way coupled with a VBD cable solver | Simulated cable insertion by an RB-Y1 robot; Samsung plans to use Newton for synthetic data to train VLA models [32] |

**ETH Zurich ANYmal-D (September 2025).** In the Newton Beta walkthrough, NVIDIA trained the ANYmal-D robot to walk on flat rigid terrain using the rsl_rl framework, with an example command running 4,096 parallel environments in headless mode. [31] The policy was then run in PhysX-based Isaac Lab as a Sim2Sim check, which the authors describe as a sanity check "to ensure a policy is not overfit to a single physics engine's specific characteristics"; a YAML mapping file handled the different joint orderings the two engines parse from the robot's USD. [31] Training used only observations available on the real robot's sensors, such as the IMU and joint encoders. With ETH Zurich's Robotic Systems Lab, the policy was deployed directly to a physical ANYmal, which executed a walking gait. [31] ETH's lab also uses Newton's implicit MPM solver to simulate soil, gravel, and stones interacting with heavy earthmoving machinery. [31]

**Skild AI assembly (March 2026).** Skild AI trains reinforcement learning policies for GPU rack assembly for its industrial customers, a contact-rich task that NVIDIA calls "one of the most demanding contact-rich tasks in electronics manufacturing." [32] Skild uses Isaac Lab with the Newton backend and routes contacts through Newton's SDF-based collision and hydroelastic contact model, bypassing MuJoCo Warp's native collision and contact pipeline to get higher contact fidelity on non-convex CAD geometry of connectors and boards. [32]

**Samsung and Lightwheel cable assembly (March 2026).** Lightwheel is using Newton to generate SimReady assets tuned against real-world measurements, including for cable manipulation in Samsung manufacturing workflows. NVIDIA notes that cables "exhibit complex 1D deformable behavior, self-collision, and force-dependent shape changes that canonical solvers cannot capture accurately." [32] The demonstration shows an RB-Y1 robot inserting a refrigerator water-hose connector into its housing, simulated with MuJoCo Warp two-way coupled to a VBD cable solver through proxy bodies that mirror the robot links into the cable simulation. [32] The example was also shown in the GTC 2026 session "An Introduction to the Newton Physics Engine for Robotics." [29][32]

## What is domain randomization?

A simulator that exactly matched reality would let you train policies in simulation and deploy them. No simulator does. The reality gap, the difference between simulated and real dynamics, lighting, and sensors, is the source of most sim-to-real failures.

[Domain randomization](https://aiwiki.ai/wiki/domain_randomization) is the dominant practical fix. The technique was introduced for vision in the 2017 paper *Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World* by Tobin, Fong, Ray, Schneider, Zaremba, and Abbeel (then at OpenAI and UC Berkeley, IROS 2017). [18] The core idea, in the authors' words, is that "with enough variability in the simulator, the real world may appear to the model as just another variation." [18] The original paper trained a real-world object detector accurate to 1.5 cm using only synthetic data with random non-photorealistic textures, which the authors describe as the first successful transfer of a deep neural network trained only on simulated RGB images to the real world for robotic control. [18]

Domain randomization came into the mainstream when OpenAI's Dactyl project randomized physical properties such as friction coefficients and object appearance to train a Shadow Dexterous Hand to reorient objects, [19] and (in a 2019 result) to solve a Rubik's cube, using only simulated experience. [20] The Rubik's cube policy ran on the real hand after training only in simulation, and the team tested its robustness with perturbations it had not been trained on, including poking the cube with a plush giraffe. [20] That result relied on automatic domain randomization (ADR), which "automatically generates a distribution over randomized environments of ever-increasing difficulty." [20]

A short list of the things people typically randomize:

- visual: textures, colors, lighting, distractor objects, camera intrinsics and extrinsics, image noise
- dynamics: friction coefficients, contact stiffness, motor latency, joint backlash, payload mass
- sensing: IMU bias, encoder noise, latency, packet loss
- environment: obstacle layout, slope, surface compliance

This is one of those techniques that sounds dumb until it works. Throw a wide enough net and the policy learns features that are invariant to the things you randomized, which often happen to be the things that vary in reality.

## What is sim-to-real transfer?

[Sim-to-real](https://aiwiki.ai/wiki/sim_to_real_transfer) (sometimes written *sim2real*) is the umbrella term for getting a policy trained in simulation to work on real hardware. Domain randomization is one piece of it, but the broader toolkit also includes:

- **System identification.** Use real-world data to fit the simulator's parameters (mass, friction, link lengths) before training. This narrows the reality gap before randomization has to cover it.
- **Domain adaptation.** Learn a mapping (often adversarial or contrastive) between simulated and real observations so the policy sees a similar distribution at deployment.
- **Real-to-sim-to-real.** Reconstruct the real world (often with NeRF or 3D Gaussian Splatting) into a simulator, train inside it, then deploy. NVIDIA's Omniverse NuRec, for example, builds interactive driving scenes from real multi-sensor data. [53]
- **Sim-to-sim validation.** Run a policy trained in one engine inside another before touching hardware. NVIDIA argues that "a policy that can successfully transfer between simulators, like PhysX and Newton, has a much higher chance of working on a physical robot." [31]
- **Hand-eye calibration and sensor modeling.** A surprising amount of sim-to-real work is just modeling the camera, IMU, motor, and tactile sensor properly.
- **Co-training and finetuning on real data.** Pre-train in sim, finetune in real with a smaller dataset.

Quadruped locomotion is the clearest success of sim-to-real. ETH Zurich's ANYmal policies trained with thousands of parallel simulated robots transferred to the real robot, [45][31] and NVIDIA's 2024 Spot workflow trained a locomotion policy in Isaac Lab that NVIDIA says transfers zero-shot to the physical robot. [62] Manipulation is harder, partly because contact and friction are harder to simulate accurately, but RoboCasa, ManiSkill3, the Stanford BEHAVIOR programs, and contact-model work in engines such as Newton are pushing the state of the art. [14][15][32]

## What is differentiable simulation?

A differentiable simulator can take gradients of physical quantities with respect to actions, parameters, or initial conditions. That lets gradient-based optimization replace some uses of reinforcement learning. Instead of sampling thousands of trajectories, you backpropagate through the dynamics directly.

| Engine | Differentiable? | Notes |
| --- | --- | --- |
| Brax | Yes (JAX) | First-class differentiability; widely used for JAX-based control research |
| MuJoCo MJX | Yes (MJX-JAX) | Differentiability is mostly supported in the JAX implementation (MJX-JAX) but not in MJX-Warp [34] |
| MuJoCo Warp | Not yet | Differentiability via Warp "is not yet available," per the project README [36] |
| Genesis | Partial | At launch, the MPM and tool solvers supported differentiability, with other solvers planned [9] |
| Drake | Partial | Analytical gradients in some subsystems; AutoDiff scalars |
| DiffTaichi | Yes | Research framework for differentiable physics in Taichi |
| Newton | Yes, solver-dependent | Described as differentiable; support for differentiation varies by solver [37] |

Differentiable simulation is not a clean win. Contact and friction are non-smooth, so naive gradients can be biased or noisy, and sample-based methods like PPO often still beat gradient-based methods for tasks that involve a lot of contact. Where differentiable sim has worked well is for soft-body manipulation, parameter identification, trajectory optimization, and any setting where you want to optimize across many physical parameters at once.

## What are generative simulators?

The newest entry on the simulation side is not a physics engine at all. It is a neural network that learns to produce video conditioned on actions. Train such a model on enough gameplay or robot footage, and you get something that behaves like a simulator: it lets you take an action, and it shows you what would happen next.

The headline projects:

- **GameNGen** (Valevski, Leviathan, Arar, Fruchter, August 2024). A diffusion model that simulates id Software's Doom interactively at 20 FPS on a single TPU. [23] The model is trained on recordings from an RL agent that played the game, and human raters are "only slightly better than random chance" at distinguishing short clips of the real game from the simulation, even after 5 minutes of auto-regressive generation. [23] The paper, *Diffusion Models Are Real-Time Game Engines*, was published at ICLR 2025.
- **Genie 1, 2, 3** (Google DeepMind). Genie 1 (February 2024) generated action-controllable virtual worlds, mostly 2D, from text, image, photo, or sketch prompts after training on unlabelled internet videos. [57][24] Genie 2 (December 2024) extends this to 3D, supports first-person, isometric, and third-person views, and maintains a consistent world for up to a minute. [24] Genie 3, announced as a limited research preview in August 2025, generates navigable worlds from text at 720p and 24 FPS in real time while staying consistent for a few minutes. [25] Waymo built its 2026 driving world model on Genie 3. [50]
- **NVIDIA Cosmos** ([NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos)). Announced at CES in January 2025 and expanded with a major release on March 18, 2025 that added Cosmos Reason, an open reasoning model for physical AI. [27][47] Cosmos World Foundation Models include Cosmos Predict (future-frame prediction), Cosmos Transfer (controllable generation from structured inputs such as segmentation, depth, and lidar), and Cosmos Reason. [47] NVIDIA describes Cosmos as enabling "synthetic data generation to augment training datasets, simulation to test and debug physical AI models before they're deployed in the real world, and reinforcement learning in virtual environments." [27] Early adopters named at launch included robotics companies 1X, Agility Robotics, and XPENG and AV developers Uber and Waabi. [27] Cosmos Predict 2.5 and Cosmos Transfer 2.5 followed in October 2025. [48]
- **Sora** ([Sora](https://aiwiki.ai/wiki/sora)). OpenAI's February 2024 technical report, *Video generation models as world simulators*, argued that "continued scaling of video models is a promising path towards the development of capable simulators of the physical and digital world," though Sora is primarily a video generation model. [56]

These systems are not drop-in replacements for physics engines. They have strong visual fidelity but no built-in guarantees of physical consistency, contact mechanics, or mass. What they offer is *coverage*: arbitrary scenes, arbitrary actions, no need to model the world by hand. The likely future is a combination, with classical physics simulators handling contact-rich manipulation and generative models handling visual diversity, sensor simulation, and rare scenarios, which is roughly how Waymo describes its use of a world model to generate events "that are almost impossible to capture at scale in reality." [50]

## How do world models relate to simulation in reinforcement learning?

A closely related but distinct line of work is the use of learned [world models](https://aiwiki.ai/wiki/world_model) inside RL itself. Instead of using the simulator only at training time, a world model is a neural network that the agent rolls out in its own head during training and even during deployment.

The canonical paper is Ha and Schmidhuber's *World Models* (2018), which trained a VAE plus a recurrent network on rollouts from environments including a car-racing task and VizDoom, and showed that an agent could be trained "entirely inside of its own hallucinated dream generated by its world model" and transfer the policy back to the actual environment (the dream experiment used VizDoom). [21] Hafner's Dreamer line (DreamerV1 in 2019, DreamerV2 in 2020, published at ICLR 2021, and DreamerV3 in 2023) generalized this. DreamerV3, published in *Nature* in April 2025 as *Mastering diverse control tasks through world models* (Hafner, Pasukonis, Ba, Lillicrap), outperforms specialized methods across more than 150 diverse tasks with a single configuration and was the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula. [22] Other notable model-based RL methods include MuZero, IRIS (transformer-based world models), TD-MPC2, and Sutton's Dyna lineage going back to the 1990s.

The boundary between "a simulator" and "a world model" has gotten blurry. Genie 2, GameNGen, and Cosmos behave like simulators (you can take actions and observe results) but are trained from data rather than coded. World models in Dreamer behave like policies' internal simulators. The unifying view is that anything that lets an agent ask "what happens if I do X?" is, functionally, a simulator.

## What are the recent milestones (2024-2026)?

The field has moved fast. Markers from the past two years, as of September 25, 2026:

| Date | Milestone |
| --- | --- |
| December 4, 2024 | Google DeepMind introduces Genie 2, a foundation world model for action-controllable 3D environments. [24] |
| December 18, 2024 | Genesis open-sources its multi-physics Python engine, reporting 43 million FPS for a Franka manipulation scene on one RTX 4090. [44][9] |
| January 6, 2025 | NVIDIA introduces Cosmos world foundation models at CES 2025. [27] |
| March 18, 2025 | At GTC, NVIDIA, Google DeepMind, and Disney Research announce Newton, an open-source physics engine built on NVIDIA Warp, and Google DeepMind introduces MuJoCo-Warp. Disney's Star Wars-inspired BDX droids join Jensen Huang on stage, and NVIDIA says Disney Research will be among the first to use Newton. [26][28] The same day NVIDIA announces Isaac GR00T N1, which it billed as the world's first open humanoid robot foundation model, [28] and a major Cosmos release with Cosmos Reason. [47] |
| August 5, 2025 | Google DeepMind unveils Genie 3, a real-time text-to-world model running at 720p and 24 FPS. [25] |
| August 8, 2025 | MuJoCo 3.3.5 adds Warp as a backend implementation for MJX. [35] Isaac Sim 5.0.0 reaches general availability on GitHub. [41] |
| September 29, 2025 | Disney Research, Google DeepMind, and NVIDIA contribute Newton to the Linux Foundation; NVIDIA releases Newton Beta and demonstrates an ANYmal-D policy trained with Newton and deployed on a real robot with ETH Zurich. [38][31] |
| October 2025 | Cosmos Predict 2.5 and Cosmos Transfer 2.5 released. [48] |
| February 6, 2026 | Waymo introduces the Waymo World Model, built on Genie 3. [50] |
| February 12, 2026 | MuJoCo 3.5.0 marks MuJoCo Warp's official release. [35] |
| February 18, 2026 | RoboCasa365 v1.0: 365 tasks, 2,500+ kitchen scenes, 2,200+ hours of demonstration data. [54] |
| March 16, 2026 | Newton 1.0 announced as generally available at GTC 2026, with the Kamino solver, VBD deformables, SDF and hydroelastic contact, and Skild AI and Samsung/Lightwheel industrial examples. [32] |
| March 17, 2026 | Isaac Lab 3.0 Beta released. [40] |
| May 27, 2026 | Genesis AI releases Genesis World 1.0, the renamed Genesis platform. [43][42] |
| June 4, 2026 | Isaac Sim 6.0.0 released. [41] |
| September 10, 2026 | Newton v1.6.0 and Isaac Sim 6.1.0 released. [39][41] |
| September 16, 2026 | Isaac Lab 3.0 Early Access, built for Isaac Sim 6.1 and Newton 1.5.2, with general availability targeted for the end of October 2026. [40] |
| September 22, 2026 | MuJoCo and MuJoCo Warp 3.14.0 released. [35][36] |
| September 22, 2026 | NVIDIA publishes its "Physics Simulation for Robotics" use-case page covering the MuJoCo, MJWarp, Warp, Newton, Isaac Sim, and Isaac Lab stack; the NVIDIA Robotics account promoted it on X on September 24. [29][30] |

A reasonable read: the line between a physics simulator, a generative video model, and a robot foundation model is collapsing. Classical physics increasingly handles contact, while learned models handle visual diversity and rare events.

## What are the limitations and open problems?

Simulation is not a solved problem. The honest list of what still goes wrong:

- **Contact and friction.** Rigid-contact mechanics are still surprisingly hard. Different engines disagree on the same scene, and small differences in friction coefficients, restitution, and integration step size produce qualitatively different policies. Sim-to-sim checks across engines are one response. [31]
- **Deformable objects.** Cloth, cables, food, and human bodies remain harder than rigid bodies. NVIDIA notes that such tasks "may require a specialized solver or coupled physics approach rather than a rigid-body-only model." [29] GPU solvers for them are maturing, including Newton's VBD and implicit MPM solvers and Genesis's FEM, MPM, and particle solvers. [32][42]
- **Photorealism vs. speed.** Photoreal rendering and high-throughput physics still pull in opposite directions on hardware budgets. Isaac Lab 3.0's answer is to make them separate choices: Isaac Sim with PhysX and RTX rendering for sensor-rich workflows, or headless Newton physics for high throughput. [37][40]
- **Sensor realism.** Cameras, LiDARs, IMUs, and tactile sensors are imperfectly modeled. Tactile simulation is an active area, with work such as hydroelastic contact data in Newton and Peking University's Taccel solver. [32][31]
- **The reality gap is irreducible without real data.** Even with heavy domain randomization, a policy that has never seen a real robot usually underperforms one that has been finetuned on it. Real-to-sim-to-real pipelines and continual learning are partial answers.
- **Generative simulators have no physics guarantees.** Learned world models such as Genie or GameNGen do not enforce conservation laws or contact mechanics by construction, so objects can behave implausibly. Whether that is a fixable training problem or a structural limitation is open.
- **Benchmark fragility.** Simulation benchmarks tend to overfit. A policy that wins on RoboCasa is not necessarily a generalist robot, just like a model that wins on ImageNet was not always a strong vision system.
- **Vendor benchmarks.** Many headline speed numbers (Genesis's 43 million FPS, MuJoCo Warp's multiples over MJX) come from the developers themselves. NVIDIA's own guidance is to benchmark your own workload. [29]

None of this means simulation is going away. The opposite. Every part of the modern stack assumes it. But the gap between "works in the simulator" and "works in the real world" remains the central engineering problem of robotics and embodied AI.

## See also

- [MuJoCo](https://aiwiki.ai/wiki/mujoco)
- [NVIDIA Newton](https://aiwiki.ai/wiki/nvidia_newton)
- [NVIDIA Warp](https://aiwiki.ai/wiki/nvidia_warp)
- [NVIDIA Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim)
- [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab)
- [PyBullet](https://aiwiki.ai/wiki/pybullet)
- [Gazebo](https://aiwiki.ai/wiki/gazebo_simulator)
- [Brax](https://aiwiki.ai/wiki/brax)
- [Genesis](https://aiwiki.ai/wiki/genesis_simulator)
- [Drake](https://aiwiki.ai/wiki/drake)
- [Webots](https://aiwiki.ai/wiki/webots)
- [CARLA](https://aiwiki.ai/wiki/carla)
- [AirSim](https://aiwiki.ai/wiki/airsim)
- [Habitat](https://aiwiki.ai/wiki/ai_habitat)
- [AI2-THOR](https://aiwiki.ai/wiki/ai2_thor)
- [Domain randomization](https://aiwiki.ai/wiki/domain_randomization)
- [Sim-to-real transfer](https://aiwiki.ai/wiki/sim_to_real_transfer)
- [World model](https://aiwiki.ai/wiki/world_model)
- [Dreamer](https://aiwiki.ai/wiki/dreamer)
- [Genie](https://aiwiki.ai/wiki/genie)
- [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos)
- [Reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning)
- [Robotics](https://aiwiki.ai/wiki/robotics)
- [Robot learning](https://aiwiki.ai/wiki/robot_learning)
- [Physical AI](https://aiwiki.ai/wiki/physical_ai)

## References

1. Todorov, Erez, and Tassa. *MuJoCo: A physics engine for model-based control.* IROS 2012. doi:10.1109/IROS.2012.6386109
2. DeepMind. *Open-sourcing MuJoCo.* Blog post, May 23, 2022. https://deepmind.google/discover/blog/open-sourcing-mujoco/
3. Coumans, Erwin. Bullet Physics SDK and PyBullet. https://pybullet.org/ ; https://github.com/bulletphysics/bullet3
4. Open Robotics. Gazebo and Ignition Gazebo. https://gazebosim.org/ ; Wikipedia: *Gazebo (simulator)*.
5. Makoviychuk et al. *Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning.* NeurIPS Datasets and Benchmarks, 2021. arXiv:2108.10470.
6. NVIDIA. Isaac Lab Documentation (including migration guides from IsaacGymEnvs, OmniIsaacGymEnvs, and Orbit). https://isaac-sim.github.io/IsaacLab/
7. Freeman, Frey, Raichuk, Girgin, Mordatch, and Bachem. *Brax: A Differentiable Physics Engine for Large Scale Rigid Body Simulation.* arXiv:2106.13281, NeurIPS 2021 Datasets and Benchmarks Track.
8. Tedrake, Russ, and the Drake Development Team. *Drake: Model-Based Design and Verification for Robotics.* https://drake.mit.edu/ ; credits: https://drake.mit.edu/credits.html
9. Genesis Authors. *Genesis: A Generative and Universal Physics Engine for Robotics and Beyond.* December 2024. https://github.com/Genesis-Embodied-AI/Genesis (README as of December 30, 2024) ; project page: https://genesis-embodied-ai.github.io/ (December 2024 snapshot: https://web.archive.org/web/20241225000000/https://genesis-embodied-ai.github.io/)
10. Cyberbotics. *Webots.* Open-sourced December 2018. https://cyberbotics.com/ ; https://github.com/cyberbotics/webots
11. Savva et al. *Habitat: A Platform for Embodied AI Research.* ICCV 2019.
12. Puig et al. *Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots.* Meta AI, October 2023. arXiv:2310.13724.
13. Kolve et al. *AI2-THOR: An Interactive 3D Environment for Visual AI.* arXiv:1712.05474.
14. Nasiriany et al. *RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots.* RSS 2024. arXiv:2406.02523
15. Tao et al. *ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI.* arXiv:2410.00425, 2024.
16. Dosovitskiy, Ros, Codevilla, Lopez, and Koltun. *CARLA: An Open Urban Driving Simulator.* CoRL 2017. arXiv:1711.03938.
17. Microsoft Research. *AirSim.* https://github.com/microsoft/AirSim
18. Tobin, Fong, Ray, Schneider, Zaremba, and Abbeel. *Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World.* IROS 2017. arXiv:1703.06907.
19. OpenAI. *Learning Dexterity.* Blog post, July 2018. https://openai.com/index/learning-dexterity/ ; paper: OpenAI et al., *Learning Dexterous In-Hand Manipulation*, arXiv:1808.00177.
20. OpenAI et al. *Solving Rubik's Cube with a Robot Hand.* 2019. arXiv:1910.07113. https://openai.com/index/solving-rubiks-cube/
21. Ha and Schmidhuber. *World Models.* arXiv:1803.10122, 2018. Interactive version: https://worldmodels.github.io/
22. Hafner, Pasukonis, Ba, and Lillicrap. *Mastering Diverse Domains through World Models.* arXiv:2301.04104, 2023; published as *Mastering diverse control tasks through world models*, Nature, April 2, 2025. https://www.nature.com/articles/s41586-025-08744-2
23. Valevski, Leviathan, Arar, and Fruchter. *Diffusion Models Are Real-Time Game Engines (GameNGen).* arXiv:2408.14837, ICLR 2025. https://gamengen.github.io/
24. Google DeepMind. *Genie 2: A Large-Scale Foundation World Model.* December 4, 2024. https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/
25. Google DeepMind. *Genie 3: A New Frontier for World Models.* August 5, 2025. https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/
26. NVIDIA. *Announcing Newton, an Open-Source Physics Engine for Robotics Simulation.* March 18, 2025. https://developer.nvidia.com/blog/announcing-newton-an-open-source-physics-engine-for-robotics-simulation/
27. NVIDIA Blog (Ming-Yu Liu). *Cosmos World Foundation Models Openly Available to Physical AI Developers.* January 6, 2025. https://blogs.nvidia.com/blog/cosmos-world-foundation-models/
28. NVIDIA. *NVIDIA Announces Isaac GR00T N1, the World's First Open Humanoid Robot Foundation Model, and Simulation Frameworks.* March 18, 2025. arXiv:2503.14734. https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks
29. NVIDIA. *Physics Simulation for Robotics* (use case page). Accessed September 25, 2026. https://www.nvidia.com/en-us/use-cases/physics-simulation-for-robotics/
30. NVIDIA Robotics (@NVIDIARobotics). "Physics simulation gives robots a place to learn before they enter the physical world." X post, September 24, 2026. https://x.com/NVIDIARobotics/status/2103137379569590394
31. NVIDIA Technical Blog (Mohajerani, Shanmuganathan, Heiden, Chen, Daviet, Rakhsha, Richard). *Train a Quadruped Locomotion Policy and Simulate Cloth Manipulation with NVIDIA Isaac Lab and Newton.* September 29, 2025. https://developer.nvidia.com/blog/train-a-quadruped-locomotion-policy-and-simulate-cloth-manipulation-with-nvidia-isaac-lab-and-newton/
32. NVIDIA Technical Blog (Reist, Zamora Mora, Chang, Chadha, Mohajerani). *Newton Adds Contact-Rich Manipulation and Locomotion Capabilities for Industrial Robotics.* March 16, 2026. https://developer.nvidia.com/blog/newton-adds-contact-rich-manipulation-and-locomotion-capabilities-for-industrial-robotics/
33. MuJoCo Documentation. *MuJoCo Warp (MJWarp).* https://mujoco.readthedocs.io/en/latest/mjwarp/index.html
34. MuJoCo Documentation. *MuJoCo XLA (MJX).* https://mujoco.readthedocs.io/en/latest/mjx.html
35. MuJoCo Documentation. *Changelog* (versions 3.3.5 through 3.14.0). https://mujoco.readthedocs.io/en/latest/changelog.html
36. google-deepmind/mujoco_warp. *MuJoCo Warp (MJWarp)*, GitHub README and releases (v3.14.0, September 22, 2026). https://github.com/google-deepmind/mujoco_warp
37. Nuñez Cano, Patel, Arranz, Ben Horin, Lo, and Chadha (NVIDIA). *The State of Simulation for Physical AI: An Overview.* Hugging Face community blog, July 21, 2026. https://huggingface.co/blog/nvidia/state-of-simulation-for-physical-ai
38. Linux Foundation. *Linux Foundation Announces Contribution of Newton by Disney Research, Google DeepMind and NVIDIA to Accelerate Open Robot Learning.* September 29, 2025. https://www.linuxfoundation.org/press/linux-foundation-announces-contribution-of-newton-by-disney-research-google-deepmind-and-nvidia-to-accelerate-open-robot-learning
39. newton-physics/newton. GitHub releases (v1.0.0 through v1.6.0, September 10, 2026). https://github.com/newton-physics/newton/releases
40. isaac-sim/IsaacLab. *Isaac Lab 3.0 Early Access* (v3.0.0-EA), GitHub release, September 16, 2026; and release history including v3.0.0-beta (March 17, 2026). https://github.com/isaac-sim/IsaacLab/releases/tag/v3.0.0-EA
41. isaac-sim/IsaacSim. GitHub repository, LICENSE, and releases (5.0.0, August 8, 2025; 6.0.0, June 4, 2026; 6.1.0, September 10, 2026). https://github.com/isaac-sim/IsaacSim
42. Genesis-Embodied-AI/genesis-world. GitHub README and releases (v1.0.0, May 27, 2026). https://github.com/Genesis-Embodied-AI/genesis-world
43. Genesis AI. *The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward.* May 27, 2026. https://www.genesis.ai/blog/the-role-of-simulation-in-scalable-robotics-genesis-world-10-and-the-path-forward
44. Zhou Xian (@zhou_xian_). Genesis announcement thread, X post, December 18, 2024. https://x.com/zhou_xian_/status/1869511650782658846
45. Rudin, Hoeller, Reist, and Hutter. *Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning.* CoRL 2021. arXiv:2109.11978.
46. NVIDIA Developer. *Isaac Gym - Now Deprecated.* https://developer.nvidia.com/isaac-gym
47. NVIDIA Newsroom. *NVIDIA Announces Major Release of Cosmos World Foundation Models and Physical AI Data Tools.* March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-announces-major-release-of-cosmos-world-foundation-models-and-physical-ai-data-tools
48. NVIDIA (Hugging Face blog). *Cosmos Predict 2.5 & Transfer 2.5: Evolving the World Foundation Models for Physical AI.* October 28, 2025. https://huggingface.co/blog/nvidia/cosmos-predict-and-transfer2-5
49. Waymo. *Simulation City: Introducing Waymo's most advanced simulation system yet for autonomous driving.* July 6, 2021. https://waymo.com/blog/2021/07/simulation-city/
50. Waymo. *The Waymo World Model: A New Frontier For Autonomous Driving Simulation.* February 6, 2026. https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/
51. Microsoft. *AirSim announcement: This repository will be archived in the coming year* (README notice added July 2022, now kept in project_airsim.md). https://github.com/microsoft/AirSim/blob/main/project_airsim.md
52. LG Electronics. *Sunsetting SVL Simulator.* https://github.com/lgsvl/simulator
53. NVIDIA Developer. *Autonomous Vehicle Simulation.* https://developer.nvidia.com/drive/simulation
54. RoboCasa project site (RoboCasa365 v1.0 release notes, February 18, 2026). https://robocasa.ai/
55. Ai2. *MolmoSpaces, an open ecosystem for embodied AI.* https://allenai.org/blog/molmospaces
56. OpenAI. *Video generation models as world simulators.* February 15, 2024. https://openai.com/index/video-generation-models-as-world-simulators/
57. Bruce et al. *Genie: Generative Interactive Environments.* arXiv:2402.15391, February 2024.
58. Wikipedia. *Bullet (software).* https://en.wikipedia.org/wiki/Bullet_(software)
59. Gazebo. *About Gazebo.* https://gazebosim.org/about ; Gazebo Classic end-of-life notice: https://classic.gazebosim.org/
60. google-deepmind/mujoco_playground. *MuJoCo Playground*, GitHub README. https://github.com/google-deepmind/mujoco_playground
61. DeepMind. *Opening up a physics simulator for robotics.* October 18, 2021. https://deepmind.google/discover/blog/opening-up-a-physics-simulator-for-robotics/
62. NVIDIA Technical Blog. *Closing the Sim-to-Real Gap: Training Spot Quadruped Locomotion with NVIDIA Isaac Lab.* June 17, 2024. https://developer.nvidia.com/blog/closing-the-sim-to-real-gap-training-spot-quadruped-locomotion-with-nvidia-isaac-lab/