> Working notes for *Open-Source Humanoid Robotics Landscape (Oct 2026)*, compiled Oct 2, 2026. Not fact-checked line by line: where these notes and the report disagree, trust the report. See README.md. # Open-source humanoid robotics, simulation slice: simulators, physics engines, training frameworks and environments (status as of 2026-10-02) ## How to read this report **Source tags** - **[GH]**: GitHub repo page fetched today with WebFetch. - **[GI]**: GitHits registry/repo index. This covers PyPI metadata plus the linked repo's stars/forks; download counts were refreshed 2026-10-02. - **[GI-file]**: a repository file read through the GitHits index (a commit-pinned snapshot). - **[web]**: another web page fetched today. - **[headline]**: a search-result title only; the page was not opened. - **"(self-reported)"**: the number was published by the project's own team. **Fetch limitations** - The fetch proxy refused these pages with HTTP 429 (rate limited). I did not retry them: - github.com/google-deepmind/mujoco - github.com/google-deepmind/mujoco/releases - github.com/google-deepmind/mujoco_warp - github.com/google-deepmind/mujoco_menagerie - github.com/google-deepmind/mujoco_playground - github.com/Genesis-Embodied-AI/genesis-world - pypi.org/project/mujoco, pypi.org/project/mujoco-warp, pypi.org/project/newton - NVIDIA blog post "Newton adds contact-rich manipulation and locomotion…" - Dataconomy article on Newton 1.0 - Stone Tao's Substack post on Genesis - Because github.com was throttled, I took star counts from GitHits where possible. - Star counts for **Genesis, MuJoCo Menagerie and MuJoCo Playground could not be fetched today**. - This session's shared WebSearch quota (200 calls) ran out partway through. --- ## 0. The 2025–2026 consolidation and what it means **1. Newton became the shared GPU physics core.** - Newton is a Linux Foundation project under Apache-2.0, started by NVIDIA, Google DeepMind and Disney Research. - Releases: first PyPI build 0.1.3 on 2025-09-26, **1.0.0 on 2026-03-10**, then monthly minor releases up to **1.6.0 on 2026-09-10** [GI]. - Its primary backend is **MuJoCo Warp**. MuJoCo Warp was "officially released" in MuJoCo 3.5.0 (2026-02-12) and its version numbers now track MuJoCo's (3.14.0, 2026-09-22) [GI-file]. - Newton's Technical Steering Committee includes: - Yuval Tassa (Google DeepMind) - Miles Macklin (NVIDIA) - Vassilios Tsounis and Moritz Baecher (Disney Research) - Drake developers Michael Sherman and Joe Masterjohn (Toyota Research Institute) **2. NVIDIA's stack no longer requires Isaac Sim ("kit-less").** - Isaac Lab 3.0 went from beta on 2026-03-17 to Early Access on 2026-09-16 [GH]. It supports several physics backends: PhysX, Newton/MuJoCo Warp, and OVPhysX. - In Newton mode it runs without Isaac Sim [GI-file]. - The PhysX SDK and the Omniverse PhysX extensions are BSD-3 on GitHub. A standalone `ovphysx` pip wheel exists (0.6.3, 2026-09-16) [GI]. - Isaac Sim's source has been Apache-2.0 on GitHub since 5.0 (GA tag 2025-08-08). It still needs the proprietary Omniverse Kit SDK and NVIDIA assets [GI-file]. **3. The MuJoCo ecosystem collapsed onto MJX and MuJoCo Warp.** - Brax's own physics is deprecated; only `brax/training` is maintained [GI-file]. - MuJoCo Playground can run on either MJX or MuJoCo Warp. - mjlab puts Isaac Lab's manager-based API on top of MuJoCo Warp. - MJX itself can call MuJoCo Warp as a backend (since 3.4.0). **4. One VC-backed independent remains: Genesis World.** - Genesis World 1.0 shipped on PyPI on 2026-05-27 [GI]. - It is backed by Genesis AI ($105M seed, July 2025) [web]. - It runs on CUDA, ROCm, Metal and Vulkan via its own compiler, Quadrants, a June-2025 fork of Taichi. **5. Several platforms were retired or moved to legacy.** - Habitat: Meta stopped active development beyond v0.3.4. - Isaac Gym: NVIDIA calls it "legacy … no longer supported." - legged_gym: migrated to Isaac Lab in 2024. - Brax: physics deprecated. - Cosmos-Predict2.5: limited maintenance since Cosmos 3. **6. Renderers are converging on 3D Gaussian splats.** - Isaac Sim 6.0 NuRec, the MuJoCo Warp batch renderer, Newton's Gaussian BVH and Genesis's Nyx renderer all render splats. - So do GaussGym, DISCOVERSE and GS-Playground. **7. Asset formats are being pushed toward OpenUSD, but MJCF still dominates the MuJoCo family.** - Newton imports USD and is aligning with the AOUSD deformable-schema proposals. - The `newton-physics` org publishes `mujoco-usd-converter` and `urdf-usd-converter`, which are still small (75 and 72 stars) [GI]. **What this means for a simulation, environment or data startup** - Rigid and articulated physics is becoming a free, well-funded commodity. Building another solver is a poor bet. - Defensible value is in the layers above the solver: - sim-ready assets supplied in both USD and MJCF, tuned per solver; - procedurally varied scenes and task suites; - automated demonstrations; - actuator and system-ID calibration (MuJoCo shipped a sysID toolbox and actuator delays in 3.5.0); - evaluation infrastructure that shows simulation results predict real-world results; - real-to-sim pipelines that produce collision- and physics-ready digital twins, not just splats. - Target the backend-agnostic layers: Isaac Lab 3 multi-backend, Newton and MuJoCo Warp. - Watch the licence traps: - the Omniverse Kit SDK; - cuRobo, which is proprietary and required by Isaac Lab Mimic; - Omniverse-licensed pip wheels. --- ## 1. Physics engines and core simulators ### MuJoCo (with MJX): reference engine for contact-rich robotics - **Status:** ACTIVE. - 3.14.0 released 2026-09-22. - Monthly minor releases since 3.5.0 (2026-02-12). - Last push 2026-10-01 [GI]. - **Who:** Google DeepMind (Yuval Tassa and team). - **Links:** - GitHub google-deepmind/mujoco - mujoco.org - mujoco.readthedocs.io - **GitHub stats:** 15k stars, 1.7k forks, 309 open issues [GI]. Latest release 3.14.0 (2026-09-22). - **License:** Apache-2.0. - **What it is:** - Generalized-coordinate dynamics with a convex soft-contact solver (pyramidal or elliptic friction). - Supports tendons, muscles, "flex" deformables and a rich set of sensors. - MJCF format plus the mjSpec model-editing API. - Python bindings, a WebAssembly build (`wasm/` in the repo), and an in-browser viewer at live.mujoco.org. - MJX is a JAX port (vmap/pmap on GPU or TPU). Since 3.4.0, MJX's pmap also works with MuJoCo Warp. - **2025–26 changes** [GI-file changelog]: - **3.4.0 (2025-12-05):** sleeping islands (preliminary) and quadratic flex. - **3.5.0 (2026-02-12):** - MuJoCo Warp officially released. - Python **System Identification toolbox**. - **Actuator and sensor delays** via history buffers. - Implicit integration for flex. - Rangefinder sensors attachable to cameras. - **3.14.0 (2026-09-22):** - Experimental IPC mode for penetration-free (frictionless) flex contact. - Multi-input actuators (pid, dcmotor, orientation). - Archive resource providers for `.mjz`/`.zip` files. - **Adoption:** - PyPI `mujoco`: **4.0M downloads/month** [GI]. - dm_control (4.7k stars) is still released in lockstep (1.0.47 on 2026-09-22). - It underpins Menagerie, Playground, mjlab, robosuite/RoboCasa, LocoMuJoCo, MyoSuite and the Gymnasium MuJoCo environments. - It is the usual "sim2sim" check before deploying Isaac Lab policies; for example, unitree_rl_lab goes Isaac Lab → MuJoCo → real G1 [GI-file]. - **Benchmarks:** CPU MuJoCo is tuned for per-step latency. The docs recommend MuJoCo Warp for throughput. - **Trade-offs:** - Best accuracy and determinism, minimal dependencies, and real-time MPC on CPU. - Basic OpenGL rendering; no RTX-class sensors. - USD only through a converter. - **Sources:** - github.com/google-deepmind/mujoco/blob/main/doc/changelog.rst and doc/mjwarp/index.rst [GI-file] - PyPI metadata [GI] ### MuJoCo Warp (MJWarp): GPU MuJoCo for NVIDIA hardware - **Status:** ACTIVE. - "Officially released" 2026-02-12. - PyPI releases: 0.0.1/0.0.2 (2026-01-08), then 3.5.0 (2026-02-13) through 3.14.0 (2026-09-22), at a monthly cadence [GI]. - **Who:** Google DeepMind and NVIDIA, "as part of the Newton project" (README). mjlab credits Erik Frey and Taylor Howell. - **Links:** - google-deepmind/mujoco_warp - docs at mujoco.readthedocs.io/en/latest/mjwarp - **GitHub stats:** 1.3k stars, 194 forks [GI]. **383k downloads/month** [GI]. - **License:** Apache-2.0. - **What it is:** - The MuJoCo pipeline rewritten in NVIDIA Warp; usable as a drop-in replacement for MuJoCo. - A ray-traced **batch renderer** supporting meshes, textures, heightfields, flex objects, **Gaussian splats**, per-camera intrinsics, lighting and shadows. - Reachable from three ecosystems: JAX (via MJX), PyTorch (via Newton/Isaac Lab), and mjlab. - Not yet supported: - the IMPLICITFAST midpoint feature; - the PGS and noslip solvers; - plugin actuators and sensors. - Flex is experimental, and Warp-based differentiability is "not yet available" (issue #500) [GI-file README]. - **Adoption:** - Primary backend of Newton. - Isaac Lab 3.0's Newton backend. - mjlab. - MuJoCo Playground (`--impl warp`). - Amazon FAR's holosoma. - RoboVerse (via mjlab). - **Benchmarks:** - Nightly benchmarks cover G1 flat and heightfield, MyoArm, ALOHA (clutter, pot, SDF, cloth), three humanoids and cloth. The page is JavaScript-rendered and I could not extract numbers from it. - Renderer claim: "millions of frames per second" (self-reported; no conditions given). - The best concrete public numbers are in Isaac Lab's benchmark data (see the Isaac Lab entry). - **Trade-offs** (official docs): - "Scales better than MJX on nearly every workload." - A single step is slower than CPU MuJoCo. - It scales "not as well as MuJoCo for single large kinematic trees … beyond \~60 DoFs remains an active priority." - That matters for humanoids with dexterous hands: a 29-DoF G1 plus two 16–24-DoF hands is well over 60 DoF. - Fast simulation requires an NVIDIA GPU; CPU works for debugging. - **Sources:** - github.com/google-deepmind/mujoco_warp/blob/main/README.md [GI-file] - doc/mjwarp/index.rst [GI-file] - google-deepmind.github.io/mujoco_warp/nightly/ [web, page had no extractable data] ### MuJoCo Menagerie: curated MJCF robot models - **Status:** likely ACTIVE. - The indexed README includes recent hands and arms (Sharpa Wave, Tetheria Aero Hand Open, Franka FR3 v2, Trossen WXAI, i2rt YAM). - Newton 1.6 (Sep 2026) re-converted "eight structured USD menagerie robots." - Commit dates are unverified (GitHub page 429). - **Who:** Google DeepMind, with vendor and community contributions. - **GitHub stats:** not fetched today. - **License:** set per model (Apache-2.0, BSD-3, BSD-2, MIT, BSD-3-Clear). - **What it is:** the gallery lists **71 models** [GI-file]: - **11 humanoids:** - Unitree H1 (19 DoF) and G1 (29) - Robotis OP3 (20) - PAL TALOS (44) - Booster T1 (23) - ToddlerBot 2XC and 2XM (44 each) - PNDbotics Adam_lite (25) - Apptronik Apollo (32) - Berkeley Humanoid (12) - Fourier N1 (23) - **Dexterous hands:** - Allegro V3 (16) - Shadow E3M5 (24) - Shadow DEX-EE (12) - LEAP (16) - Sharpa Wave (22) - Tetheria Aero Hand Open (16) - **Grippers:** Robotiq 2F-85, UMI. - **Other:** about 25 arms, mobile manipulators (Stretch, TIAGo, RBY1, TidyBot), quadrupeds, MS-Human-700 and flybody. - **Adoption:** Playground downloads Menagerie assets automatically. Newton and Isaac Lab Newton presets use USD conversions of it. - **Trade-offs:** high-quality, but small and MJCF-first. The licence differs per model, which matters if you redistribute. - **Sources:** github.com/google-deepmind/mujoco_menagerie/blob/main/README.md [GI-file]. ### MuJoCo Playground: GPU robot-learning environments with sim-to-real recipes - **Status:** on the ACTIVE/SLOWING boundary. PyPI `playground` 0.1.0 (2026-01-08) and 0.2.0 (2026-03-16) [GI]; commit dates unverified. - **Who:** Google DeepMind and UC Berkeley (Zakka, Tabanpour, Liao, … Sreenath, Sferrazza, Tassa, Abbeel). - **Paper:** arXiv 2502.08844 (2025-02-12), RSS 2025. - **GitHub stats:** not fetched (429). - **License:** Apache-2.0. - **What it is:** - dm_control classic tasks. - Quadruped and biped/humanoid locomotion. - Non-prehensile and dexterous manipulation. - Vision tasks via the MuJoCo Warp batch renderer. - Trains with Brax PPO (JAX) or RSL-RL; switch to MuJoCo Warp with `--impl warp`. - **Benchmarks and sim2real** (paper abstract): "Train policies in minutes on a single GPU." Zero-shot sim-to-real from both state and pixels on quadrupeds, humanoids, dexterous hands and arms. - **Pain point:** the README warns that JAX defaults to TF32 on Ampere/Ada GPUs, which hurts RL reproducibility. Set `JAX_DEFAULT_MATMUL_PRECISION=highest`. - **Trade-offs:** fastest route to locomotion or hand sim2real in JAX. mjlab (PyTorch) and Isaac Lab (RTX sensors, USD) suit other stacks. - **Sources:** - arxiv.org/abs/2502.08844 [web] - PyPI `playground` description [GI] ### Newton: Linux Foundation GPU physics engine for robotics (Warp-based) - **Status:** ACTIVE. - 0.1.3 (2025-09-26), 0.2.0 (2026-02-25), **1.0.0 (2026-03-10)**. - 1.1.0 (Apr 13), 1.2.0 (May 12), 1.3.0 (Jun 11), 1.4.0 (Jul 16), 1.5.0 (Aug 11). - **1.6.0 (2026-09-10)** [GI-file CHANGELOG]. - The Linux Foundation announced Newton 1.0 general availability around GTC in March 2026 [headline]. - **Who:** - Started by Disney Research, Google DeepMind and NVIDIA. - **Maintainers:** Miles Macklin, Saran Tunyasuvunakool, Vassilios Tsounis, Eric Heiden, Jan Carius, Eric Shi, @nvtw and Alain Denzler. - **TSC:** Macklin; Tsounis (co-chair); Michael Sherman; Joe Masterjohn; Yuval Tassa; Moritz Baecher; Louis Lian; Mohammad Mohajerani (co-chair); Frank Chen [GI-file newton-governance/CONTRIBUTORS.md]. - **Links:** - newton-physics/newton - newton-physics.github.io/newton - **GitHub stats:** **5.6k stars, 671 forks** [GH]. Latest release 1.6.0 (2026-09-10). - **License:** Apache-2.0 for code, CC-BY-4.0 for docs. - **What it is:** generalizes Warp's deprecated `warp.sim`. One Model/State/Control API with interchangeable solvers [GI-file docs/solvers]: - **Generalized coordinates:** SolverMuJoCo (MuJoCo Warp) and SolverFeatherstone. - **Maximal coordinates:** SolverSemiImplicit, SolverXPBD, and SolverKamino (experimental; kinematic loops, hard frictional contact). - SolverVBD (experimental; cloth, soft and rigid bodies in one solver). - SolverStyle3D (cloth). - SolverImplicitMPM (particles and granular). - **Features:** - SDF and mesh collision. - Hydroelastic contact (experimental). - Speculative contacts (1.6). - Gaussian-splat BVH. - Tiled ray-traced camera sensor. - Import from OpenUSD (including AOUSD deformable proposals), MJCF and URDF. - Differential-IK and operational-space controllers; DC-motor actuator models. - **Platforms:** Python 3.10+, Linux (x86-64/aarch64) and Windows. Requires an NVIDIA GPU (Maxwell or newer, driver 545+). **macOS runs CPU only.** - **Adoption:** - Isaac Lab 3.0 Newton backend. - ProtoMotions 3 (pinned to Newton 1.0) [GI-file]. - The Linux Foundation's 1.0 announcement lists "stable articulated mechanism simulation, hydroelastic contact modeling, deformable body simulation (cables, cloth, rubber)…" [headline: x.com/linuxfoundation]. - **Benchmarks:** see the Isaac Lab table: Newton/MuJoCo Warp delivers about 1.5–2.1× PhysX training throughput with less VRAM (NVIDIA-run). Third-party claims such as "475× faster" (byteiota) are unverified. - **Trade-offs:** - The broadest multi-solver open engine, with backing from industry and the Linux Foundation. - The API is changing quickly; 1.6 deprecates cable APIs and changes USD deformable defaults, for example. - Speed requires NVIDIA hardware. - Several solvers are still experimental. - **Sources:** - github.com/newton-physics/newton [GH] - README.md, CHANGELOG.md and docs/solvers/index.rst [GI-file] - newton-governance CONTRIBUTORS.md and CONTRIBUTING.md [GI-file] ### NVIDIA Isaac Sim (with PhysX/OVPhysX): Omniverse/USD robot simulator with RTX sensors - **Status:** ACTIVE. - GitHub tags [GH]: - v5.0.0 GA (2025-08-08) - v5.1.0 GA (2025-10-21) - v6.0.0-dev (2025-12-19) - v6.0.0-dev2 (2026-03-16, GTC'26 early developer release) - **v6.0.0 GA (2026-06-04)** - v6.0.1 (2026-06-22) - 6.1: - PyPI `isaacsim` 6.1.0.0 (2026-09-09) [GI]. - Forum post "Isaac Sim 6.1 General Availability" [headline]. - The repo VERSION file read 6.1.0-rc.26 when indexed. - An open issue complains the 6.1.0 release notes are missing [headline]. - **Who:** NVIDIA. - **GitHub stats:** isaac-sim/IsaacSim **\~4.0k stars, 534 forks** [GH]. PhysX repo NVIDIA-Omniverse/PhysX: 4.7k stars, 671 forks [GI]. - **License:** partially open. - The repo code is **Apache-2.0**. - Building or using it "requires additional components licensed under other terms … Omniverse Kit SDK, as well as 3D models and textures" (the Isaac Sim Additional Software and Materials License) [GI-file LICENSE]. - PyPI metadata for the `isaacsim` and `isaaclab` wheels says "NVIDIA Proprietary Software" [GI]. - PhysX SDK and Omni PhysX extensions are BSD-3. The `ovphysx` wheel's license field is "LicenseRef-NVIDIA-Omniverse" [GI]. - **What it is:** - Imports URDF, MJCF and CAD; native USD. - GPU PhysX; RTX multi-sensor rendering. - Synthetic data generation, ROS 2 bridge, digital twins. - **6.0 added NuRec 3D Gaussian splat scenes**, GPU nested rigid bodies and expanded XR [web]. - Requirements: Python 3.12; Ubuntu 22.04/24.04 (GCC 11 only), Windows 11, or aarch64 (DGX Spark). - **Minimum GPU: RTX 4080 (workstation) or A40 (datacenter).** - The build needs internet access to Kit extension servers [GI-file README]. - **Adoption:** BEHAVIOR-1K/OmniGibson, Isaac Lab, robosuite (photoreal rendering option), and Arena-based benchmarks. - **Trade-offs:** - Highest-fidelity sensors and USD/Omniverse integration. - Heavy installation, high GPU floor, proprietary Kit dependency, and frequent breaking upgrades. - **Sources:** - github.com/isaac-sim/IsaacSim and /releases [GH] - README.md, LICENSE and VERSION [GI-file] - NVIDIA-Omniverse/PhysX README [GI-file] - radiancefields.com article on Isaac Sim 6.0 NuRec [web] ### Isaac Lab (including 3.0): NVIDIA's open robot-learning framework - **Status:** ACTIVE [GH releases]: - v2.3.1 (2025-12-04) - v2.3.2 (2026-02-02) - **v3.0.0-beta (2026-03-17)** - v3.0.0-beta2 (2026-06-17) - beta2.patch1 (2026-07-02) - **v3.0.0-EA (2026-09-16)**, which targets Isaac Sim 6.1 - PyPI `isaaclab` 2.3.2.post1 (2026-02-11) [GI] - **Who:** NVIDIA with many partners. The tech report (arXiv 2511.04831, 2025-11-06) has a long author list including ETH Zurich (Hutter), UT Austin (Zhu) and others. - **GitHub stats:** **8.2k stars, 3.9k forks** [GH]. - **License:** - BSD-3 for the framework; Apache-2.0 for `isaaclab_mimic`. - Full workflows (PhysX, RTX, ROS, URDF/MJCF importers) need Isaac Sim's proprietary components; kit-less Newton workflows do not. - **`isaaclab_mimic` requires cuRobo, which is proprietary** [GI-file README]. - **What it is:** a manager-based and direct RL/IL framework with 16+ robots and 30+ environments, including: - G1, H1 and Cassie locomotion; - Shadow-hand cube reorientation; - Franka and UR10 reach. - **3.0 architecture** [GI-file release notes]: - Factory-based **multi-backend physics**: `isaaclab_physx` (default) and `isaaclab_newton`, which is powered by MuJoCo Warp and supports the MuJoCo Warp, XPBD and Featherstone solvers with CUDA graphs. OVPhysX is experimental. - Pluggable renderers: Isaac RTX (needs Isaac Sim), OVRTX (kit-less RTX) and Newton Warp (kit-less). - Pluggable visualizers and a camera image-signal-processing pipeline (PPISP). - Isaac Capture teleop record/replay and policy export. - **Adoption:** - unitree_rl_lab (Isaac Lab 2.3 / Isaac Sim 5.1; Go2, H1, G1-29). - BeyondMimic (Isaac Lab 2.1). - holosoma (Isaac Sim path). - RoboTwin (Arena branch). - The 3.9k forks suggest heavy downstream customization. - **Benchmarks:** from NVIDIA's in-repo `environment-performance-develop.csv`: - Setup: RTX PRO 6000 Blackwell Server, 8,192 environments, RSL-RL training. - Metric: mean total frames per second (FPS), end to end. - Run dates: 2026-09-09, or 2026-08-28 where marked †. | Task | PhysX | Newton/MuJoCo Warp | OVPhysX | |---|---|---|---| | Velocity-Flat-G1 | 136,950 FPS (11.6 GB VRAM) | **264,026** (4.3 GB) | 176,802 (9.3 GB) | | Velocity-Flat-H1 | 220,167 | **377,693** | 295,423† | | Velocity-Rough-G1† | 112,096 | **169,352** | 152,929 | | Reorient-Cube-Shadow (direct) | 112,490 | **238,886** | 68,977 | - Caveats: NVIDIA ran these itself. Rows are flagged "dirty working tree reported." Newton rows are often flagged "coefficient of variation >25%." Contact parameters are not identical across backends. - A live dashboard is at nvidia.github.io/omniperf. - Older ManiSkill comparison (self-reported, Isaac Lab v1.2.0, RTX 4090, CartPole with 640×480 RGB, 128 environments): Isaac Lab 4,556 FPS using 15.2 GB, versus ManiSkill 10,144 FPS using 3.7 GB. - **Trade-offs:** - Default choice for humanoid sim2real; the richest sensors and assets; now able to switch backends. - 3.0 is still EA/beta with breaking changes. - Licensing is mixed (Kit, cuRobo). - **Sources:** - github.com/isaac-sim/IsaacLab and /releases [GH] - README.md, docs/source/refs/release_notes.rst and docs/source/_static/benchmarks/environment-performance-develop.csv [GI-file] - arxiv.org/abs/2511.04831 [web] ### Isaac Lab-Arena: composable benchmark authoring and policy evaluation - **Status:** ACTIVE (alpha). README v0.3 alpha targets Isaac Sim 6.0 / Isaac Lab 3.0, with the warning "Do not use this in production" [GI-file]. - **Who:** NVIDIA. - **GitHub stats:** 557 stars, 102 forks [GH]. - **License:** Apache-2.0. - **What it is:** - Environments are composed from a **scene**, an **embodiment** and a **task**. - Objects are placed by relation ("on", "next to"). - Atomic skills chain into long-horizon tasks. - **Agentic environment generation (experimental):** a natural-language description becomes a specification, matching SimReady USD assets are found, and a family of environments is built. - Perturbation sweeps over lighting, cameras and mass. - Thousands of heterogeneous environments per GPU; multi-node runs via OSMO. - Policy client-server interface for GR00T and π0.5. - Subtask predicates and sensitivity analysis. - **Adoption:** RoboTwin 2.0 maintains an IsaacLab-Arena branch. - **Trade-offs:** directly addresses scene and task generation and evaluation at scale, the same space a startup would target. NVIDIA-stack only and immature. - **Sources:** - github.com/isaac-sim/IsaacLab-Arena [GH] - README.md [GI-file] ### mjlab: Isaac Lab's API on MuJoCo Warp - **Status:** ACTIVE. 1.6.0 (2026-08-09) bumps MuJoCo/MuJoCo Warp to 3.11; last push 2026-09-23 [GI]. - **Who:** Kevin Zakka, Qiayuan Liao, Brent Yi, Louis Le Lay, Koushil Sreenath and Pieter Abbeel. Paper arXiv 2601.22074 (2026-01-29). - **GitHub stats:** **3.1k stars, 538 forks**; **54k downloads/month** [GI]. - **License:** Apache-2.0, plus BSD-3 for utilities forked from Isaac Lab. - **What it is:** - Isaac Lab's manager-based environment API on MuJoCo Warp, installable with one command. - Direct access to native MuJoCo data structures. - Reference tasks: G1 velocity tracking, motion imitation (including a BeyondMimic reproduction) and manipulation. - Multi-GPU training via torchrunx; nightly benchmarks. - **Benchmarks:** "on par or faster than Isaac Lab" (self-reported FAQ; no numbers). - **Pain points** [GI-file FAQ and releases]: - The FAQ attributes NaN crashes partly to MuJoCo Warp "still in beta." - 1.5.3 fixed "CUDA 700 crashes from mass domain randomization on certain GPU architectures." - 1.5.2 fixed multi-environment domain-randomization crashes. - Recommended GPUs: RTX 40-series or newer, L40S, H100. - **Trade-offs:** a lightweight PyTorch stack with no Omniverse. It lacks RTX sensors and USD asset tooling. - **Sources:** - github.com/mujocolab/mjlab README.md and docs/source/faq.rst [GI-file] - arxiv.org/abs/2601.22074 [web] ### Genesis World: multiphysics simulator from Genesis AI - **Status:** ACTIVE. - 1.0.0 (2026-05-27). - Roughly weekly releases since then, up to **1.4.3 (2026-09-30)** [GI]. - The repo was renamed to `genesis-world`. - **Who:** - Started as an academic project in December 2024; now "officially supported by Genesis AI." - Founders: Zhou Xian (CEO) and Théophile Gervet (President). - **$105M seed (July 2025)** led by Eclipse and Khosla Ventures, with Bpifrance, HSG, Eric Schmidt, Xavier Niel, Daniela Rus and Vladlen Koltun. - About 60 staff in Paris, California and London. - The company has gone "full stack": its own hand (with Wuji Tech), a data glove and the GENE-26.5 model. Zhou says simulation mainly speeds up **evaluation** [web TechCrunch 2026-05-06]. - Media in 2026 reported talks to raise about $500M at about a $3B valuation [headline, unverified]. - **Links:** - Genesis-Embodied-AI/genesis-world - genesis-world.readthedocs.io - Quadrants (219 stars) and Nyx renderer (`nyx-for-genesis`, 138 stars) [GI] - **GitHub stats:** stars and forks **not fetched** (GitHub page 429). - **License:** Apache-2.0. - **What it is:** - One scene and state shared by rigid, FEM, MPM, SPH, PBD, IPC (libuipc) and SAP coupling. - Three renderers: Nyx (with 3D Gaussian splats), Luisa ray tracing and Pyrender. - **Quadrants compiler** targets CUDA, AMD ROCm, Apple Metal, Vulkan, x86 and ARM64. - Imports URDF, MJCF, OBJ, GLB and USD. - Sensors include tactile, IMU, lidar, depth and contact force. - G1 and dexterous-hand benchmarks run in CI. - **Benchmarks and the 2024 controversy:** - The original README (v0.2.1/v0.3.0) claimed "**over 43 million FPS when simulating a Franka robotic arm with a single RTX 4090 (430,000× faster than real-time)**" [GI-file]. - In December 2024, ManiSkill lead Stone Tao published "How fast is the new hyped Genesis simulator?" It argued the figure came from an unrepresentative, near-trivial setup and that Genesis was not faster on comparable manipulation workloads. I could not re-open the post today (429), so its specifics are not re-verified here. - On 2025-01-09 Genesis published a benchmarking report with test scripts (`zhouxian/genesis-speed-benchmark`, which links to a Notion page I could not read). - The current README drops the 43M-FPS claim. Contributors must now show at least +5% on the CI benchmarks in `tests/benchmarks`. - Self-reported release notes: - 1.0 made the rigid solver "up to 35% faster for contact-rich scenes with \~64 DoFs." - A later 1.x release runs "100 entities / 1000 geometries … real-time at 120 FPS" on CPU. - 1.4.2 notes that Signorini contact is still "experimental due to convergence issue." - **I found no independent head-to-head benchmark of Genesis 1.x.** - **Trade-offs:** - The widest multiphysics and the only cross-vendor GPU option, plus tactile sensors. - Credibility on speed claims was damaged in 2024; the API changes quickly; the ecosystem is smaller than NVIDIA's or MuJoCo's. - **Sources:** - README.md, RELEASE.md and README at v0.2.1/v0.3.0 [GI-file] - techcrunch.com/2026/05/06/… [web] - Substack post [headline only; fetch refused] ### ManiSkill3 / SAPIEN: GPU-parallel manipulation simulation and rendering - **Status:** - **ManiSkill:** ACTIVE (3.0.1 released 2026-04-21; last push 2026-08-04) [GI]. - **SAPIEN:** SLOWING (3.0.3 released 2026-03-10) [GH][GI]. - The README now points to the github.com/mani-skill org. - **Who:** UCSD SU Lab (Hao Su) and Hillbot. Paper arXiv 2410.00425 (Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, …), RSS 2025. - **GitHub stats:** ManiSkill 3.3k stars, 550 forks [GI]; SAPIEN 793 stars, 77 forks [GH]. - **License:** ManiSkill Apache-2.0; SAPIEN MIT. - **What it is:** - PhysX-5 GPU simulation with SAPIEN's Vulkan renderer (rasterized; ray tracing is not parallelized). - **Heterogeneous parallel scenes:** a different scene in every parallel environment. - Task suites: table-top, mobile manipulation, humanoid (Unitree G1, MuJoCo humanoid; "still adding more"), dexterous. - Real-to-sim evaluation environments. - Baselines: PPO/SAC/TD-MPC2, BC/Diffusion Policy, and VLA models (Octo, RDT-1B, RT-x). - URDF and MJCF loaders; MJCF support is partial. - **Benchmarks** (self-reported): - Rendering plus simulation "10–1000× faster with 2–3× less GPU memory." - "30,000+ FPS." - CartPole comparison: see the Isaac Lab entry. - **Downstream:** RoboTwin 2.0 (2.8k stars); RoboVerse integration. - **Trade-offs:** - Easiest pip installation and very efficient visual RL for manipulation. - GPU simulation is Linux-only (CPU only on Windows). - Thin humanoid coverage; small team. - **Sources:** - github.com/haosulab/SAPIEN [GH] - ManiSkill README, docs/source/user_guide/additional_resources/performance_benchmarking.md, benchmark CSVs and custom_robots.md [GI-file] - arxiv.org/abs/2410.00425 [web] ### Drake: model-based robotics toolbox with high-fidelity contact - **Status:** ACTIVE. - Monthly releases; PyPI `drake` 1.55.0 (2026-07-15), 1.56.0 (2026-08-14), **1.57.0 (2026-09-10)** [GI]. - The GitHub page fetch reported v1.54.0 (June 12, 2026) as latest, which conflicts with PyPI; the PyPI dates are more recent. - **Who:** Toyota Research Institute (Russ Tedrake; Sherman and Masterjohn, who also sit on Newton's TSC). - **GitHub stats:** 4.1k stars, 1.4k forks [GH]. - **License:** BSD-3. - **What it is:** - MultibodyPlant with hydroelastic contact and the SAP solver. - Parsers for URDF, SDFormat and (partially) MJCF. - Optimization and verification tools; CPU-based. - **Trade-offs:** - Best for verifying manipulation and model-based control. - Not designed for massively parallel RL; small humanoid-RL footprint. - **Sources:** github.com/RobotLocomotion/drake [GH]; PyPI [GI]. ### Gazebo (Harmonic, Ionic, Jetty): ROS-centric robot simulator - **Status:** ACTIVE [web gazebosim.org/docs/all/releases]: - **Jetty:** LTS, September 2025, end of life May 2031. GitHub tag `gz-sim10_10.0.0` dated 2025-10-14 [GH]. - **Ionic:** end of life December 2026. - **Harmonic:** LTS until May 2029. - **Fortress:** LTS until May 2027. - Next release "M" planned for March 2027 (new cadence). - "Rotary" is a nightly rolling release. - **Who:** Open Robotics / OSRF; the gz-sim maintainer address is at Intrinsic. - **GitHub stats:** gz-sim 1.4k stars, 427 forks [GH]. - **License:** Apache-2.0. - **What it is:** - Physics via gz-physics plugins: **DART, Bullet, Bullet-Featherstone and TPE**. - OGRE 2 rendering; SDFormat (URDF via conversion); ROS 2 integration. - **Trade-offs:** standard for ROS system and integration testing and sensor simulation. Not GPU-parallel; rarely used for humanoid RL. - **Sources:** - github.com/gazebosim/gz-sim [GH] - gz-sim and gz-physics READMEs [GI-file] - releases page [web] ### Webots: desktop simulator for education and industry - **Status:** SLOWING. - The last tagged release is **R2025a** (GitHub, 2025-02-04; changelog says 2025-01-31) [GH]. - An unreleased R2025b changelog in the repo adds Ubuntu 26.04 support, which implies 2026 commits; I did not verify commit dates. - **Who:** Cyberbotics. - **GitHub stats:** 4.5k stars, 2k forks [GH]. - **License:** Apache-2.0. - **What it is:** ODE-based physics, PROTO model format, URDF import, and many bundled robots, including humanoids such as the KHR-3HV. - **Trade-offs:** easy to use and good for teaching. CPU-only and not suited to large-scale RL. - **Sources:** github.com/cyberbotics/webots [GH]; docs/reference/changelog-r2025.md [GI-file]. ### Project Chrono: multiphysics engine for vehicles, terrain, FEA and granular materials - **Status:** ACTIVE. **10.0.0 released 2026-03-27**; 9.0.0 was 2024-05-20 [GI-file CHANGELOG]. - **Who:** University of Wisconsin–Madison SBEL and partners. - **GitHub stats:** 3.1k stars, 638 forks [GH]. - **License:** BSD-3. - **What it is:** - Multibody dynamics, FEA, DEM granular materials, fluid-structure interaction and vehicles. - Templates for ROS, FMI2 and vehicle co-simulation. - PyChrono via conda. - **Trade-offs:** strong for deformable terrain and off-road vehicles. Niche for humanoids or GPU RL. - **Sources:** github.com/projectchrono/chrono [GH]; CHANGELOG.md [GI-file]. ### PyBullet: legacy Python physics simulator - **Status:** SLOWING, close to dormant. Last release pybullet 3.2.7 (2025-01-30); last push to bullet3 2025-10-22 [GI]. - **Who:** Erwin Coumans and the community. - **GitHub stats:** bullet3 14k stars, 3.0k forks [GI]. - **License:** zlib. - **What it is:** CPU rigid-body simulation with URDF, SDF and MJCF (partial) support. It still underpins older benchmarks (GenSim/Ravens) and Habitat. - **Trade-offs:** use only for legacy reproduction. - **Sources:** PyPI [GI]. ### Brax: JAX RL library (physics deprecated) - **Status:** ACTIVE, but README: "Only `brax/training` is actively being maintained as of 0.13.0." It points users to MuJoCo Playground for environments and MJX/MuJoCo Warp for physics, and says Brax may become purely an RL library. Latest 0.14.2 (2026-03-15); last push 2026-09-29 [GI]. - **GitHub stats:** 3.2k stars, 356 forks [GI]. - **License:** Apache-2.0. - **Sources:** google/brax README.md [GI-file]. ### Isaac Gym: deprecated GPU RL simulator - **Status:** DISCONTINUED. NVIDIA: "This is legacy software. Developers may download and continue to use it, but it is no longer supported." It recommends Isaac Lab instead. Preview 4 is the final version [web]. - legged_gym announced its migration to Isaac Lab on 2024-01-09 [GI-file]. - Still used by holosoma, ProtoMotions, HumanoidVerse and GaussGym. - **License:** proprietary NVIDIA preview license (not open source). - **Sources:** developer.nvidia.com/isaac-gym [web]; legged_gym and IsaacGymEnvs READMEs [GI-file]. --- ## 2. Benchmark and environment suites ### RoboCasa and robosuite: MuJoCo-based household manipulation at scale - **Status:** ACTIVE. - **RoboCasa365 v1.0 (2026-02-18)**; leaderboard opened 2026-04-06; v1.0.1 (2026-05-12); per-frame subtask annotations added 2026-07-07 [GI-file]. - robosuite 1.5.2 (2025-12-24); last push 2026-07-11 [GI]. - **Who:** UT Austin (Yuke Zhu's lab) and collaborators. RoboCasa365 paper: arXiv 2603.04356, ICLR 2026 [headline]. - **GitHub stats:** RoboCasa 1.5k stars, 222 forks [GH]; robosuite 2.6k stars, 777 forks [GI]. - **License:** MIT code; **CC BY 4.0 assets and datasets**. robosuite is MIT [GI-file]. - **What it is:** - RoboCasa365 has 365 tasks created "with the guidance of LLMs." - **2,500+ kitchen scenes and 3,200+ objects**, including AI-generated objects. - **600+ hours of human demonstrations plus 1,600+ hours of robot data** produced with automated trajectory tools. - Benchmarks Diffusion Policy, π and GR00T. - Assets are about a 10 GB download. - robosuite 1.5 (2024-10-31) added humanoids, composite and whole-body controllers, Apple Vision Pro teleoperation and Isaac Sim photoreal rendering. - **Adoption:** NVIDIA released Cosmos-Predict2.5 "robot/policy" checkpoints for RoboCasa and LIBERO on 2026-02-23 [GI-file]. - **Trade-offs:** the richest open household content. CPU MuJoCo, so not GPU-parallel; dependencies are fragile ("use the master branch!"). - **Sources:** github.com/robocasa/robocasa [GH]; README.md [GI-file]; robosuite LICENSE and v1.5.0 notes [GI]. ### BEHAVIOR-1K / OmniGibson: 1,000-activity household benchmark on Isaac Sim - **Status:** ACTIVE. - OmniGibson 3.9.3 in the monorepo; last push 2026-09-29 [GI]. - A 2026 BEHAVIOR Challenge has demonstration datasets on Hugging Face [headline]. - PyPI `omnigibson` is stale at 1.1.1 (2024). - **Who:** Stanford (Li Fei-Fei, Jiajun Wu and colleagues). - **GitHub stats:** 1.7k stars, 250 forks [GH]. - **License:** MIT for code. - **What it is:** 1,000 everyday activities, Isaac Sim/PhysX with fluids, cloth and object states, BDDL task definitions, and LeRobot-format datasets. - **Trade-offs:** the deepest long-horizon household semantics. Heavy (Isaac Sim) and slow compared with GPU RL simulators. - **Sources:** github.com/StanfordVL/BEHAVIOR-1K [GH]; README and OmniGibson/setup.py [GI-file]. ### Habitat 3 (habitat-sim / habitat-lab): embodied AI navigation and rearrangement - **Status:** DISCONTINUED by Meta. Both READMEs: "**Beyond v0.3.4 this project is no longer receiving official active development or maintenance by Meta internal teams.**" Last PyPI publish and push 2026-05-07 [GI]. - **GitHub stats:** habitat-sim **3.8k stars, 548 forks**, not archived [GH]; habitat-lab 3.1k stars, 691 forks [GI]. - **License:** MIT. - **What it is:** Bullet physics; URDF robots; HM3D, HSSD and ReplicaCAD scenes; humanoid avatars. - **Benchmarks:** "Fetch … in ReplicaCAD at over 8,000 SPS (1 RGBD 128×128 observation + 1/30 s dynamics per step)" (self-reported). - **Trade-offs:** fast navigation research, but the platform risk is now high. - **Sources:** github.com/facebookresearch/habitat-sim [GH]; READMEs [GI-file]. ### RoboVerse / MetaSim: one API across many simulators, plus a dataset - **Status:** ACTIVE. **1.0.0b0 and MetaSim 0.2.0 released 2026-05-31** (CHANGELOG); design docs dated May 2026 [GI-file]. - **Who:** academic consortium (BIGAI among contributors). Paper arXiv 2504.18904, RSS 2025. - **GitHub stats:** 1.9k stars, 167 forks [GH]. - **License:** Apache-2.0. - **What it is:** - MetaSim handlers for Isaac Lab/Isaac Sim, Isaac Gym, MuJoCo, SAPIEN, PyBullet, Genesis and CoppeliaSim, plus Blender rendering. - Integrations with LIBERO, ManiSkill, mjlab, robosuite, RoboTwin and EmbodiedGen. - **Trade-offs:** useful for migrating assets and tasks between simulators. Lowest-common-denominator abstractions. - **Sources:** github.com/RoboVerseOrg/RoboVerse [GH]; README and CHANGELOG [GI-file]. ### Gymnasium-Robotics: Farama MuJoCo robotics environments - **Status:** ACTIVE (maintenance). 1.4.2 (2026-01-02); last push 2026-09-07 [GI]. - **GitHub stats:** 1.9k stars, 150 forks. - **License:** MIT. - **What it is:** Fetch, Shadow Hand, Adroit, Franka Kitchen, mazes and MaMuJoCo. CPU-based; legacy-style benchmarks. - **Sources:** PyPI [GI]. --- ## 3. Procedural and generative environment tools - **Infinigen (Indoors / Articulated "Infinigen-Sim"):** ACTIVE. - **v2.0.0a1 (2026-07-15), "first alpha release of 2.0 overhaul."** - Princeton VL; **7.1k stars, 608 forks**; BSD-3 [GH]. - Procedural Blender generation of nature scenes, indoor scenes (CVPR 2024) and **sim-ready articulated assets** (arXiv 2505.10755). - The most credible open procedural asset generator. Its outputs need physics tuning per simulator. - **Holodeck (allenai/Holodeck):** likely DORMANT. - No releases; data version 2023_09_23; commit dates unverified. - 572 stars, 65 forks; Apache-2.0 [GH]. - LLM-guided scene generation for AI2-THOR (CVPR 2024, arXiv 2312.09067). AI2-THOR's last PyPI release was 5.0.0, three years ago [GI]. - **RoboGen (Genesis-Embodied-AI/RoboGen):** likely DORMANT. 1.2k stars, 108 forks; Apache-2.0; no releases [GH]. Generative robot-agent task, scene and training pipeline. - **GenSim (liruiw/GenSim):** likely DORMANT. 350 stars, 25 forks; MIT; no releases [GH]. GPT-4 writes PyBullet/Ravens task code (arXiv 2310.01361). - **Productized successors in 2026:** - RoboCasa365's LLM-guided tasks and AI-generated objects. - Isaac Lab-Arena's agentic environment generation (experimental). - RoboVerse's EmbodiedGen integration. - The academic generators of 2023–24 mostly were not maintained. --- ## 4. Real-to-sim and digital twins (Gaussian splatting) - **Isaac Sim 6.0 NuRec** [web radiancefields.com, 2026-06-07]: - Reconstructed real scenes load as USD and render through the Fabric Scene Delegate (multi-GPU, light interaction with meshes, MaterialX). - **3DGRUT** exports USDZ. - Physics requires **separately paired collision proxies**. - **GaussGym** (arXiv 2510.15352, 2025-10-17; Escontrela, Kerr, Allshire, Frey, Duan, Sferrazza, Abbeel): - Uses 3DGS as the renderer inside **Isaac Gym**. - "**>100,000 simulation steps per second on consumer GPUs**" (self-reported). - Scenes come from iPhone scans, GrandTour, ARKit and **Veo-generated video**. - Shows sim-to-real for locomotion from pixels. Repository stars not fetched. - **DISCOVERSE** (TATP-233/DISCOVERSE): MuJoCo plus 3DGS for Real2Sim2Real; IROS 2025; arXiv 2507.21981; 518 stars, 65 forks; MIT [GH]. - **GS-Playground** (RSS 2026; Yufei Jia et al.): custom parallel physics engine plus 3DGS; "**10⁴ FPS at 640×480**" (self-reported) [web roboticsproceedings]. - **Engine-native support:** - MuJoCo Warp's batch renderer renders splats. - Newton exposes `Gaussian.bvh`. - Genesis Nyx has a 3DGS example. - **Gap:** splats carry appearance only, not mass, friction or articulation. Turning them into physics-ready twins is still manual. --- ## 5. World models used as simulators - **NVIDIA Cosmos:** - **Cosmos 3** released May 2026 (Nano and Super; Edge 4B in July 2026). - One Mixture-of-Transformers model with two interfaces: a "Reasoner" (text and vision to text) and a "Generator" (text, vision, sound and action to vision, sound and action). - Includes action-conditioned forward dynamics and policies. - **Code and models under OpenMDW-1.1** [GI-file]. - Repo NVIDIA/Cosmos: **11.7k stars, 854 forks** [GH]. - NVIDIA claims first place among open models on several physical-AI benchmarks (self-reported) [web eWeek]. - The Cosmos Coalition's founding members (per eWeek) include Agile Robots, Generalist, Skild AI and Runway. - Cosmos-Predict2.5 (2025-10-06; 2B/14B; robot action-conditioned models; DreamGen post-training) and Transfer2.5 are now in "limited maintenance" [GI-file]. - **1X World Model:** - The production model shipped 2026-01-13 for the NEO robot. TechCrunch reports no information on weights or licence, so treat it as closed [web]. - The open **1xgpt** challenge repo (561 stars, 48 forks, Apache-2.0 [GH]) provides: - more than 100 hours of tokenized EVE robot video with actions; - a GENIE-style baseline and a MAGVIT2 tokenizer; - compression and sampling challenges ($10k each); - an evaluation (policy-ranking) challenge marked "upcoming" [GI-file]. - Challenge technical report: arXiv 2510.07092 [headline]. - **Closed comparators:** - **Veo world simulator** for Gemini Robotics policy evaluation (arXiv 2512.10675, 2025-12-11): validated against "1,600+ real-world evaluations of eight Gemini Robotics policy checkpoints and five tasks." It predicts relative policy ranking in nominal and out-of-distribution conditions. No release stated [web]. - **Genie 3:** "Project Genie" rolled out to Google AI Ultra users in the US around 2026-01-29 [headline]. No open weights. --- ## 6. Also notable (brief) - **NVIDIA Warp** (NVIDIA/warp): ACTIVE (1.17.0, 2026-08-31); 6.9k stars; Apache-2.0; 1.0M downloads/month. `warp.sim` is deprecated in favor of Newton [GI]. - **OVPhysX / PhysX** (NVIDIA-Omniverse/PhysX): ACTIVE (ovphysx 0.6.3, 2026-09-16; PhysX SDK 5.11); 4.7k stars [GI]. - **dm_control** (google-deepmind/dm_control): ACTIVE (1.0.47, 2026-09-22); 4.7k stars [GI]. - **LocoMuJoCo** (robfiras/loco-mujoco): ACTIVE commits (push 2026-08-22; 1.0.0 released 2025-04-17); 1.4k stars; MIT; humanoid imitation benchmark [GI]. - **MyoSuite**: ACTIVE (2.12.2, about June 2026); musculoskeletal models on MuJoCo; Apache-2.0 [GI]. - **Meta-World** (Farama-Foundation/Metaworld): ACTIVE (3.1.1, about July 2026); 1.8k stars [GI]. - **Gymnasium**: ACTIVE (1.3.0); 12k stars; 3.6M downloads/month [GI]. - **rsl_rl** (leggedrobotics/rsl_rl): ACTIVE (rsl-rl-lib 5.5.1, about September 2026); 2.8k stars; 260k downloads/month; BSD-3 [GI]. - **ProtoMotions 3** (NVlabs/ProtoMotions): ACTIVE. - 2.4k stars, 461 forks; Apache-2.0 [GH]. - Backends: Newton 1.0, Isaac Lab 3.0, Isaac Gym and MuJoCo. - "AMASS 40+ hours within 12 hours on 4 A100s" (self-reported). - **holosoma** (amazon-far/holosoma): 1.6k stars, 253 forks; Apache-2.0 [GH]. - Backends: Isaac Gym, Isaac Sim, MuJoCo Warp and MuJoCo (inference only). - G1 and Booster T1; PPO and FastSAC. - **unitree_rl_lab**: Isaac Lab 2.3 / Isaac Sim 5.1; Go2, H1 and G1-29; Apache-2.0 [GI-file]; stars not fetched. - **BeyondMimic** (HybridRobotics/whole_body_tracking): Isaac Lab 2.1; MIT; arXiv 2508.08241; reproduced in mjlab [GI-file]. - **HumanoidVerse** (LeCAR-Lab/HumanoidVerse): status unverified. Pins Isaac Gym Preview 4, Isaac Sim 4.2 and Genesis 0.2.1; MIT [GI-file]. - **legged_gym**: DORMANT (migrated in 2024). - **RoboTwin 2.0** (RoboTwin-Platform/RoboTwin): ACTIVE (last update 2026-08-03); ICML 2026; 2.8k stars, 479 forks; MIT [GH]. - **RoboHive** (vikashplus/robohive): DORMANT (last PyPI release two years ago); 632 stars [GI]. - **AI2-THOR**: releases dormant (5.0.0, three years ago); 1.8k stars [GI]. - **DART**: ACTIVE (dartpy 6.19.4, about two months ago); 1.1k stars; BSD-2 [GI]. - **newton-physics/mujoco-usd-converter**: ACTIVE (0.6.1, 2026-09-29); 75 stars [GI]. - **newton-physics/urdf-usd-converter**: 0.3.3; 72 stars [GI]. - **newton-physics/newton-usd-schemas**: 0.5.0; 14 stars [GI]. --- ## 7. Comparison table | Simulator | Maintainer | License | GPU-parallel | Physics | Rendering | Formats | Humanoid / hands | Stars | Latest release | Status | |---|---|---|---|---|---|---|---|---|---|---| | MuJoCo (+MJX) | Google DeepMind | Apache-2.0 | MJX (JAX) | MuJoCo soft-contact | OpenGL | MJCF, URDF (partial) | Yes / yes | 15k | 3.14.0 (2026-09-22) | ACTIVE | | MuJoCo Warp | Google DeepMind + NVIDIA | Apache-2.0 | Yes (CUDA) | MuJoCo | Ray-traced batch, 3DGS | MJCF | Yes / yes (>60-DoF trees slow) | 1.3k | 3.14.0 (2026-09-22) | ACTIVE | | Newton | Linux Foundation (NVIDIA, GDM, Disney) | Apache-2.0 | Yes (CUDA; macOS CPU only) | MuJoCo Warp, Featherstone, XPBD, VBD, Kamino, MPM, Style3D | Warp ray-trace, 3DGS | USD, MJCF, URDF | Yes / yes | 5.6k | 1.6.0 (2026-09-10) | ACTIVE | | Isaac Sim | NVIDIA | Apache-2.0 + proprietary Kit/assets | Yes | PhysX 5 (+Newton) | RTX, NuRec 3DGS | USD, URDF, MJCF, CAD | Yes / yes | \~4.0k | 6.0.1 (GitHub); 6.1.0 (PyPI, 2026-09-09) | ACTIVE | | Isaac Lab | NVIDIA + community | BSD-3 / Apache-2.0 (+ cuRobo proprietary) | Yes | PhysX, Newton, OVPhysX | Isaac RTX, OVRTX, Newton Warp | USD (+importers) | Yes / yes | 8.2k | v3.0.0-EA (2026-09-16) | ACTIVE | | mjlab | mujocolab (Berkeley) | Apache-2.0 | Yes | MuJoCo Warp | MuJoCo Warp / viser | MJCF | G1 / limited | 3.1k | 1.6.0 (2026-08-09) | ACTIVE | | Genesis World | Genesis AI | Apache-2.0 | Yes (CUDA, ROCm, Metal, Vulkan) | Rigid, FEM, MPM, SPH, PBD, IPC, SAP | Nyx, Luisa, Pyrender | URDF, MJCF, USD, GLB, OBJ | Yes / yes + tactile | not fetched | 1.4.3 (2026-09-30) | ACTIVE | | ManiSkill3 / SAPIEN | UCSD SU Lab / Hillbot | Apache-2.0 / MIT | Yes (Linux) | PhysX 5 | Vulkan raster/RT | URDF, MJCF (partial) | G1 (few tasks) / yes | 3.3k / 793 | 3.0.1 (2026-04-21) / 3.0.3 (2026-03-10) | ACTIVE / SLOWING | | Drake | TRI | BSD-3 | No | Hydroelastic + SAP | VTK, Meshcat | URDF, SDFormat, MJCF (partial) | Limited | 4.1k | 1.57.0 (2026-09-10) | ACTIVE | | Gazebo (gz-sim) | Open Robotics / Intrinsic | Apache-2.0 | No | DART, Bullet, TPE | OGRE 2 | SDF, URDF | Limited | 1.4k | Jetty 10.0.0 (2025) | ACTIVE | | Webots | Cyberbotics | Apache-2.0 | No | ODE (fork) | OpenGL PBR | PROTO, URDF | Some | 4.5k | R2025a (2025-02) | SLOWING | | Chrono | UW–Madison | BSD-3 | Partial | Multiphysics | Irrlicht / VSG | URDF (parser) | Niche | 3.1k | 10.0.0 (2026-03-27) | ACTIVE | | PyBullet | Coumans / community | zlib | No | Bullet | TinyRenderer / OpenGL | URDF, SDF, MJCF | Legacy | 14k | 3.2.7 (2025-01-30) | SLOWING | | Brax | Google | Apache-2.0 | Yes (JAX) | Physics deprecated | – | – | – | 3.2k | 0.14.2 (2026-03-15) | ACTIVE (training only) | | Isaac Gym | NVIDIA | Proprietary | Yes | PhysX | Basic | URDF, MJCF | Yes | n/a | Preview 4 | DISCONTINUED | | RoboCasa / robosuite | UT Austin | MIT (+CC BY 4.0 assets) | No | MuJoCo | MuJoCo / Isaac | MJCF | Humanoids (robosuite 1.5) | 1.5k / 2.6k | v1.0.1 (2026-05) / 1.5.2 | ACTIVE | | BEHAVIOR-1K | Stanford | MIT (code) | Limited | PhysX via Isaac Sim | RTX | USD | Mobile manipulators | 1.7k | OmniGibson 3.9.3 | ACTIVE | | Habitat | Meta (ended) | MIT | No | Bullet | Magnum | URDF, glTF | Avatars | 3.8k | 0.3.3/0.3.4 (2026-05) | DISCONTINUED | | RoboVerse | Consortium | Apache-2.0 | Via backends | Many | Many + Blender | Many | Some | 1.9k | 1.0.0b0 (2026-05-31) | ACTIVE | | Gymnasium-Robotics | Farama | MIT | No | MuJoCo | MuJoCo | MJCF | Hands | 1.9k | 1.4.2 (2026-01-02) | ACTIVE | | Isaac Lab-Arena | NVIDIA | Apache-2.0 | Yes | Isaac Lab backends | RTX | USD | Yes | 557 | v0.3 alpha | ACTIVE | | Infinigen | Princeton | BSD-3 | n/a (generator) | – | Blender | Exports meshes/URDF/USD | Assets | 7.1k | v2.0.0a1 (2026-07-15) | ACTIVE | | Cosmos 3 | NVIDIA | OpenMDW-1.1 | n/a (world model) | Learned | Generated video | – | Policies | 11.7k | Cosmos 3 (2026-05) | ACTIVE | --- ## 8. Gaps and pain points (evidence) **1. Contact fidelity for dexterous humanoids is still immature on GPU.** - MuJoCo Warp's own docs say scaling past about 60 DoF in one connected mechanism "remains an active priority." - MuJoCo Warp lacks PGS and noslip, and has no differentiability yet. - Experimental contact features across engines: - Newton: Kamino, VBD and hydroelastic contact. - MuJoCo: IPC flex contact (frictionless). - Genesis: Signorini contact. - mjlab's FAQ ties NaN crashes partly to MuJoCo Warp being in beta; 1.5.3 fixed GPU-architecture-specific crashes from mass randomization. - **Opportunity:** validated contact parameter sets and system-ID services. MuJoCo 3.5 added a sysID toolbox and delay models, but there is no shared calibration data. **2. Benchmarks are mostly self-reported and hard to trust.** - The Genesis 43M-FPS dispute. - ManiSkill compared itself with Isaac Lab v1.2 using a rasterizer against ray tracing. - NVIDIA's own CSV flags "dirty working tree" and run-to-run variation above 25%. - MuJoCo Warp's nightly dashboard is not machine-readable. - **Opportunity:** a neutral benchmarking and certification service measuring throughput, fidelity and sim-to-real correlation. **3. Asset and scene creation is the bottleneck, and formats are fragmented.** - Menagerie has only 71 curated models, each with its own licence. - The USD converters are new and tiny (about 72–75 stars each). - Newton 1.6 changed USD deformable defaults (thickness, mass precedence) to align with AOUSD proposals that "are not yet registered." - ManiSkill notes "not all MJCF properties are supported." - RoboCasa365's 2,500 scenes and 3,200 objects (a 10 GB download) is the largest open household set. - The research-era generators (Holodeck, RoboGen, GenSim) are dormant. Arena's agentic generation is experimental. - **Opportunity:** validated sim-ready assets in both USD and MJCF with articulation, collision and physics parameters, plus a procedural scene service. **4. Real-to-sim produces appearance, not physics.** - NuRec needs separate collision proxies. - GaussGym, DISCOVERSE and GS-Playground render splats over hand-built physics. - **Opportunity:** automated splat-to-physics twins with segmentation, articulation inference and mass/friction estimation. **5. Licensing and lock-in.** - Isaac Sim needs the proprietary Kit SDK and assets. - Isaac Lab Mimic needs cuRobo, which is proprietary. - The `isaacsim`/`isaaclab` pip wheels are labelled "NVIDIA Proprietary Software"; the `ovphysx` wheel carries an Omniverse licence. - MuJoCo Warp and Newton need NVIDIA GPUs for speed (Genesis is the cross-vendor exception). - Isaac Sim requires at least an RTX 4080 or A40. **6. Platforms keep churning.** - Habitat dropped by Meta; Isaac Gym legacy; Brax physics deprecated; Cosmos-Predict2.5 sidelined. - Repositories moved: ManiSkill to the mani-skill org, Genesis renamed to genesis-world, RoboTwin migrated. - Isaac Lab 3.0 is a breaking re-architecture. - Newton deprecates APIs every minor release. - Users bear high maintenance costs, which favours adapters and backend-agnostic environment layers. **7. Evaluation is the stated bottleneck.** - Genesis AI's CEO: "the real bottleneck … is evaluation." - Responses so far: Isaac Lab-Arena, the RoboCasa365 and RoboTwin leaderboards, the BEHAVIOR Challenge, and GDM's Veo simulator validated on 1,600+ real trials. - **Gap:** no open, standardized method for measuring how well simulation results correlate with real-world results. **8. Reproducibility pitfalls.** - Playground's TF32 warning for JAX on Ampere/Ada GPUs. - MuJoCo's IPC mode is not captured by `mj_getState`, so exact replay is not supported. - Isaac Sim builds only with GCC 11. **9. World models need action-labeled data.** - 1X's open release is 100+ hours of a single robot (EVE). - Cosmos 3 is open (OpenMDW) but post-training needs paired video and action data. - **Opportunity:** supplying action-conditioned interaction data and evaluation sets. --- ## 9. GitHub OWNER/REPO list with stars Tags as in the header; "nf" means not fetched today. - google-deepmind/mujoco: 15k [GI] - google-deepmind/mujoco_warp: 1.3k [GI] - google-deepmind/mujoco_menagerie: nf (429) - google-deepmind/mujoco_playground: nf (429) - google-deepmind/dm_control: 4.7k [GI] - newton-physics/newton: 5.6k [GH] - newton-physics/newton-governance: nf - newton-physics/mujoco-usd-converter: 75 [GI] - newton-physics/urdf-usd-converter: 72 [GI] - newton-physics/newton-usd-schemas: 14 [GI] - NVIDIA/warp: 6.9k [GI] - NVIDIA-Omniverse/PhysX: 4.7k [GI] - isaac-sim/IsaacSim: \~4.0k [GH] - isaac-sim/IsaacLab: 8.2k [GH] - isaac-sim/IsaacLab-Arena: 557 [GH] - isaac-sim/IsaacGymEnvs: nf - mujocolab/mjlab: 3.1k [GI] - Genesis-Embodied-AI/genesis-world: nf (429) - Genesis-Embodied-AI/quadrants: 219 [GI] - Genesis-Embodied-AI/nyx-for-genesis: 138 [GI] - Genesis-Embodied-AI/RoboGen: 1.2k [GH] - zhouxian/genesis-speed-benchmark: nf - haosulab/ManiSkill (now mani-skill/ManiSkill): 3.3k [GI] - haosulab/SAPIEN: 793 [GH] - RobotLocomotion/drake: 4.1k [GH] - gazebosim/gz-sim: 1.4k [GH] - gazebosim/gz-physics: nf - cyberbotics/webots: 4.5k [GH] - projectchrono/chrono: 3.1k [GH] - bulletphysics/bullet3: 14k [GI] - google/brax: 3.2k [GI] - robocasa/robocasa: 1.5k [GH] - ARISE-Initiative/robosuite: 2.6k [GI] - StanfordVL/BEHAVIOR-1K: 1.7k [GH] - facebookresearch/habitat-sim: 3.8k [GH] - facebookresearch/habitat-lab: 3.1k [GI] - RoboVerseOrg/RoboVerse: 1.9k [GH] - Farama-Foundation/Gymnasium-Robotics: 1.9k [GI] - Farama-Foundation/Gymnasium: 12k [GI] - Farama-Foundation/Metaworld: 1.8k [GI] - princeton-vl/infinigen: 7.1k [GH] - allenai/Holodeck: 572 [GH] - allenai/ai2thor: 1.8k [GI] - liruiw/GenSim: 350 [GH] - TATP-233/DISCOVERSE: 518 [GH] - NVIDIA/Cosmos: 11.7k [GH] - nvidia-cosmos/cosmos-predict2.5: nf - nvidia-cosmos/cosmos-transfer2.5: nf - 1x-technologies/1xgpt: 561 [GH] - NVlabs/ProtoMotions: 2.4k [GH] - amazon-far/holosoma: 1.6k [GH] - leggedrobotics/legged_gym: nf - leggedrobotics/rsl_rl: 2.8k [GI] - unitreerobotics/unitree_rl_lab: nf - HybridRobotics/whole_body_tracking: nf - LeCAR-Lab/HumanoidVerse: nf - RoboTwin-Platform/RoboTwin: 2.8k [GH] - robfiras/loco-mujoco: 1.4k [GI] - vikashplus/robohive: 632 [GI] - dartsim/dart: 1.1k [GI] **Other caveats** - Some GitHits "last pushed" dates may be cached. mujoco_warp, for example, showed 2026-08-13 despite later releases, so I relied on release dates there. - The Drake GitHub sidebar (v1.54.0) conflicts with PyPI (1.57.0, 2026-09-10). I used PyPI. - I did not name any company as a user unless a source stated it.