Software tables compare versions, languages, task IDs, and published assets, not hardware payload.
A version string is not a hardware specification. A paper repository does not prove that checkpoint runs on an unmentioned robot. G1, S1, and TM01 hardware numbers stay on their class pages.
The 2026-09-03 product table adds no SDK or URDF. Same-day sitemap and GitHub still name none, and do not prove LDA-1B, WAM-TTT, LATENT, or Humanoid-GPT released checkpoints run on ET1. No ET1 table here. [M2] [M3]
The Apache engine: README title “Genesis World,” previously Genesis, academic since Dec 2024. Eno and GENE-26.5 stay on announced and foundation-models pages, not a payload cell.
The platform lists one current manual version per product. G1 hardware numbers live on the mobile-manipulators page; TM01 hardware numbers belong with accessories, not here.
| Item | Kind | Current version | Host / scope |
|---|---|---|---|
| Galbot G1 Quick Start# | robot manual | v2.2.4 [M1] | Galbot G1 [M1] |
| Galbot S1 manual# | robot manual | v2.0.0 [M1] | Galbot S1 [M1] |
| TM01 teleoperation master–slave arm# | accessory manual | v1.4.0 [M1] | Galbot G1 accessory [M1] |
| Galbot SDK# | Python / C++ API | v1.9.1 [M1] [M2] | G1 and S1 [M3] |
The GitHub README is the version-compatibility source.
The hosted docs introduce galbot_sdk with Python and C++ API references. The public developer site describes five capability areas (perception, planning, control, navigation, interaction) without publishing rates or latencies for those APIs.
| Metric | GalbotSDK |
|---|---|
| Latest SDK# | V1.9.1 [M3] |
| Release date# | 2026-07-03 [M3] |
| Compatible robot environment# | V1.17 [M3] |
| Languages# | Python and C++ [M3] |
| Python# | 3.8–3.14 [M3] |
| Host OS (stated)# | Ubuntu 20–24 [M3] |
| Stated robot models# | G1 / S1 [M3] |
| Package name# | galbot_sdk [M2] |
| Download# | GitHub GalaxyGeneralRobotics/GalbotSDK [M2] |
| API surface (site copy)# | perception, planning, control, navigation, interaction; high-level built-in algorithms and low-level joint / chassis / sensor access [M4] |
Prices and SLA figures are copied from Galbot’s public developer site. They are commercial terms, not hardware specs. Currency is CNY as printed; tax and procurement status are unpublished.
| Metric | FREE | PRO | ENTERPRISE |
|---|---|---|---|
| List price# | ¥0, permanent [M4] | ¥50,000 / robot / year [M4] | ¥200,000 / robot / year [M4] |
| 问问小盖 quota# | 5 / day [M4] | 100 / day [M4] | unlimited (site copy) [M4] |
| SLA P0 first response# | — | ≤4 h [M4] | ≤2 h [M4] |
| SLA P1 first response# | — | ≤12 h [M4] | ≤6 h [M4] |
| Stated inclusions# | GalbotSDK, docs, digital models, community Q&A [M4] | training camps, ticket system, remote debug, online sim, on-robot deploy [M4] | dedicated engineer, on-site debug, first-deploy escort (1 week on site) [M4] |
These repositories are Apache-2.0 robot descriptions. They prove publication of URDF / MJCF / USD, not payload. Full G1 and S1 hardware sheets stay on the mobile-manipulators page.
| Metric | Galbot G1 (Golf description) | Galbot S1 |
|---|---|---|
| Formats# | URDF, MJCF, USD [M5] | URDF, MJCF, USD [M6] |
| License# | Apache-2.0 [M5] | Apache-2.0 [M6] |
| Preset URDFs# | galbot_one_golf.urdf; galbot_one_golf_fixed_base.urdf [M5] | — |
| MJCF base variants# | wheeled (galbot_one_golf.xml), fixed base, planar base [M5] | — |
| USD entry# | usd/galbot_one_golf.usda [M5] | — |
| Platform listing# | G1 Description URDF · MJCF · USD [M1] | S1 Description URDF · MJCF · USD [M1] |
Task IDs are copied from the Isaac Lab environments page.
NVIDIA’s table text says “left arm of a Galbot humanoid robot.” That wording is quoted; it is not this landscape’s class for Galbot G1. These tasks do not publish Galbot payload, mass, or runtime.
| Task ID | Arm | End-effector | Mode | What the docs state | Workflow |
|---|---|---|---|---|---|
| Isaac-Stack-Cube-Galbot-Left-Arm-Gripper-RmpFlow-v0# | left [M7] | gripper [M7] | train [M7] | stack three cubes (blue, red, green) with the left arm; NVIDIA copy calls Galbot a humanoid [M7] | Manager Based [M7] |
| Isaac-Stack-Cube-Galbot-Right-Arm-Suction-RmpFlow-v0# | right [M7] | suction [M7] | train [M7] | right-arm suction cube stack; listed on the same environment page [M7] | Manager Based [M7] |
| Isaac-Stack-Cube-Galbot-Left-Arm-Gripper-Visuomotor-v0# | left [M7] | gripper [M7] | train [M7] | left-arm gripper visuomotor variant; play env also listed [M7] | Manager Based [M7] |
| Isaac-Stack-Cube-Galbot-Left-Arm-Gripper-Visuomotor-Play-v0# | left [M7] | gripper [M7] | play [M7] | play counterpart of the visuomotor task [M7] | Manager Based [M7] |
DeepMind’s Menagerie is a third-party MJCF collection. Galbot publishes G1/S1 descriptions separately; they are absent here. Sharpa Wave is an end-effector model, not a Galbot SKU.
| Model in Menagerie | Directory | Listed DoF | License on the list | Status |
|---|---|---|---|---|
| Unitree G1# | unitree_g1/ [M8] | 29 [M8] | BSD-3-Clause [M8] | listed under humanoids [M8] |
| Sharpa Wave# | sharpa_wave/ [M8] | 22 [M8] | Apache-2.0 [M8] | listed under end-effectors [M8] |
| Galbot (any)# | — | — | — | not listed; no directory in the repository [M8] |
GalaxyGeneralRobotics hosts these repositories with a stated Unitree G1 target. Real-robot result numbers stay on the research page.
| Metric | Humanoid-GPT | LATENT |
|---|---|---|
| Stated target# | Unitree G1, 29 DoF whole-body [M9] | Unitree G1 (G1TrackingGeneral; motion path UnitreeG1) [M10] |
| Simulator in README# | MuJoCo-MJX [M9] | MuJoCo [M10] |
| License# | Apache-2.0 [M9] | — |
| Control / tracking note# | real-robot deploy docs under deploy/ for Unitree G1 [M9] | default preprocess frequency 50 Hz [M10] |
| Paper# | arXiv:2606.03985 [R1] | arXiv:2603.12686 [R2] |
Native feasibility review: pinned Newton and Genesis examples, runtime requirements and remaining qualification work.
The README publishes no MuJoCo / Isaac Gym speedup factor; that cell stays empty.
| Metric | Genesis / genesis-world |
|---|---|
| Current README title# | Genesis World [M11] |
| Prior name / start# | Genesis; academic project since Dec 2024 [M11] |
| License# | Apache-2.0 [M11] |
| Language / install# | Python >=3.10,<3.14; PyPI package genesis-world [M11] |
| Quadrants backends named in the README# | CUDA, AMD ROCm, Apple Metal, Vulkan, x86, ARM64 [M11] |
| Company support statement# | “now officially supported by Genesis AI” [M11] |
| Docs / sibling repo# | genesis-world.readthedocs.io; Genesis-Embodied-AI/genesis-world [M12] |
| Throughput vs MuJoCo / Isaac Gym# | — |
Genesis AI’s 2026 product blog is company narrative, not a third-party benchmark or a substitute for the Apache README.
Genesis World 1.0 is presented as the simulation half of Genesis AI’s stack, next to GENE-26.5 and Eno.
Named pieces: Nyx, Quadrants, a unified rigid/deformable physics platform, and a simulation interface. No license, parameter count, or independent speedup table is published there. [P1]
Eno status: announced systems. GENE-26.5 policy and company details: foundation models.
Throughput versus genesis-world is unpublished here, so that row stays empty.
SAPIEN and Drake have license files but the GitHub license API returned no SPDX id; the cells say so. ManiSkill’s live repository is mani-skill/ManiSkill, not haosulab/ManiSkill.
| Metric | MuJoCo | SAPIEN | ManiSkill | Drake |
|---|---|---|---|---|
| Kind# | Physics engine [M16] | Robotics simulation framework [M17] | GPU parallel simulation and benchmark [M18] | Model-based design toolbox [M19] |
| Language# | C / C++ with Python bindings [M16] | C++ / Python [M17] | Python [M18] | C++ / Python [M19] |
| License (GitHub API)# | Apache-2.0 [M16] | See repository; API SPDX empty [M17] | Apache-2.0 [M18] | See repository; API SPDX empty [M19] |
| Org# | google-deepmind [M16] | haosulab [M17] | mani-skill [M18] | RobotLocomotion [M19] |
| Throughput vs genesis-world# | — | — | — | — |
A public SDK is not the same thing as a public robot description, and a simulation repository is not necessarily a real-robot deployment path.
This map keeps those surfaces separate. It compares repository coverage, not API quality, timing, safety, or hardware capability.
| Publisher / target | Region | Runtime / control API | Middleware | Simulation / training | Deployment path | Description source | License boundary |
|---|---|---|---|---|---|---|---|
| Booster Robotics — public repositories# | China [M20] | booster_robotics_sdk [M20] | booster_robotics_sdk_ros2 message / service interface [M20] | booster_gym; booster_train (Isaac Lab) [M20] | booster_deploy; simulation and real-robot policy path [M20] | booster_assets: K1 / T1 / T2 URDF / MJCF + motion CSV [M20] | per repository; assets BSD-3-Clause [M20] |
| LimX Dynamics — public SDK / humanoid repositories# | China [M21] | limxsdk-lowlevel, dependency-free C++11 / Python [M21] | ROS and ROS 2 deployment repositories [M21] | humanoid-mujoco-sim; humanoid-rl-isaaclab [M21] | humanoid-rl-deploy-ros2 with ONNX inference [M21] | humanoid-description (HU_D03 / HU_D04); tron2-robot-description (TRON2A variants) [M21] | named repositories Apache-2.0; target mappings remain repo-specific [M21] |
| Leju Robotics — Kuavo configurations# | China [M22] | controllers and hardware interfaces in kuavo-ros-opensource [M22] | ROS launch, messages, services, and version/config selection [M22] | MuJoCo, Gazebo, and Isaac Sim resources [M22] | real/sim launch paths in one monorepository [M22] | versioned xacro / URDF asset packages [M22] | no repository-root license declared [M22] |
| Hello Robot — Stretch 4 / SE4# | United States [M23] | stretch4_body Python API [M23] | stretch4_ros2 (Jazzy branch) [M23] | stretch4_mujoco; examples and perception repositories [M23] | body API + ROS 2 packages; web teleoperation also public [M23] | stretch4_urdf; composable xacro exports configured URDF [M23] | per repository; MuJoCo Clear BSD, examples MIT [M23] |
| UC Berkeley — Berkeley Humanoid Lite# | United States [M24] | open hardware and control code [M24] | project scripts / interfaces; no vendor SDK layer claimed | MJCF and USD assets; locomotion code [M24] | research build and real-robot documentation [M24] | separate assets repository: URDF / MJCF / USD [M24] | MIT code; CC BY-SA 4.0 assets [M24] |
| Stanford — ToddlerBot# | United States [M25] | runtime and robot-control code in the main repository [M25] | project-native interfaces; no vendor SDK layer claimed | URDF / MJCF assemblies and policy pipelines [M25] | sim and real-robot code in one research repository [M25] | multiple actuator / gripper assemblies under descriptions/ [M25] | MIT code; CC BY-NC-SA 4.0 design/assets [M25] |
These are deployment contracts, not a model leaderboard.
The columns are different artifact kinds—a foundation-model platform, an on-device VLA, and an open robot-learning library—but the rows ask the same operational questions: where inference runs, what crosses the boundary, what is available, and what the publisher explicitly excludes.
| Operational field | NVIDIA GR00T 1.7 | Gemini Robotics On-Device 2 | LeRobot asynchronous inference |
|---|---|---|---|
| Artifact kind# | open reference platform plus foundation model [M26] | vision-language-action model based on on-device Gemma models [M27] | open-source end-to-end robot-learning library [R3] |
| Execution topology# | Jetson Thor is the named real-time inference and control target [M26] | general-purpose manipulation runs on local devices [M27] | action planning may run on a separate machine and in parallel with low-level control [R3] |
| Input / output contract# | onboard video, language command, and joint positions → chunks of relative joint motions [M26] | text, images, and numerical proprioception → numerical robot actions [M27] | shared motor middleware, dataset streaming, policy inference, and control execution [R3] |
| Published data path# | real teleoperation, synthetic data, and internet-scale human video [M26] | images, text, robot sensor data, and action data; dataset details are not published [M27] | LeRobotDataset multimodal schema with remote streaming; 16K+ datasets and 2.2K+ contributors as of September 2025 [R3] |
| Distribution / access# | GR00T 1.7 weights and reference code in Early Access [M26] | select Trusted Testers only [M27] | open-source library and paper; public datasets and model implementations are supported [R3] |
| Stated scope# | general-purpose humanoid reasoning and manipulation, adapted by post-training to an embodiment [M26] | standing bi-arm manipulation across several robot types [M27] | low-cost manipulators, humanoid arms, hands, and mobile-manipulation platforms through extensible middleware [R3] |
| Published boundary# | Early Access excludes production deployment with commercial support and a stable validated feature set [M26] | high-DoF generalization is limited; mobile-platform and whole-body risks are outside the evaluated scope [M27] | a reusable research/development stack, not a vendor certification that every policy supports every connected robot [R3] |
Interoperability is an operational relationship, not a chassis specification.
VDA 5050 defines cross-manufacturer exchange with a master control system; Open-RMF coordinates fleets and shared building infrastructure. A robot appearing elsewhere on this site is not marked compatible unless its own evidence establishes that edge.
| Operational field | VDA 5050 3.0 | Open-RMF |
|---|---|---|
| Artifact kind# | open communication interface for mobile robots and master control [M28] | free, open-source, modular interoperability middleware [M29] |
| Coordination scope# | heterogeneous mobile robots from different manufacturers in one logistics environment [M28] | multiple robot fleets plus shared physical infrastructure [M29] |
| Planning / control boundary# | zones carry movement rules; autonomous robots plan locally and share paths while the master retains traffic influence [M28] | fleet sharing and coordination are provided without claiming one universal low-level robot controller [M29] |
| Infrastructure scope# | mobile-robot traffic, restricted and one-way zones, authorization areas, and power-saving actions [M28] | doors, elevators, and building-management systems alongside robot fleets [M29] |
| Distribution / governance# | version 3.0 published open source after public consultation [M28] | open source; managed by the Open Source Robotics Alliance since 2024 [M29] |
| Robot support proven on this site# | — | — |
These rows are software surfaces, not robot spec sheets.
Hardware numbers for Booster T1, Booster K1, Reachy 2, and Fourier GR-2 sit on their class pages. Fourier N1 hardware numbers are on the humanoids page.
| Field | Booster Robotics SDK | Fourier N1 SDK | Reachy 2 hardware / SDK docs |
|---|---|---|---|
| Public surface# | Booster open-source page + booster_robotics_sdk [M13] | fourier-grx-N1 docs and GitHub [M14] | Pollen hardware guide + SDK docs [M15] |
| Robots named on that surface# | T1 and K1 [M13] | N1 [M14] | Reachy 2 [M15] |
| What this landscape uses it for# | T1 hardware table; K1 hardware table [M13] | Software listing only — no N1 hardware sheet [M14] | Reachy 2 hardware table [M15] |