SimReady assets for
humanoid robots
Atlas, Figure 02, 1X NEO, Apptronik Apollo, Unitree H1 and G1, Tesla Optimus — Rigyd assets are calibrated for the whole-body contact-rich physics humanoid foundation models need, and ship to Isaac Lab, GR00T, MuJoCo MJX, and Genesis without re-authoring.
The problem
Why existing workflows fall short.
Whole-body contact is the hardest physics
A modern humanoid carries 25–30 DoF, two feet, two multi-finger hands, and a torso all making contact with the environment simultaneously. Friction, restitution, and inertials have to be right per contact site or the policy learns to exploit simulator artefacts and fails real-world transfer — the canonical sim-to-real gap for whole-body control.
Foundation models need scene diversity, not curated rooms
GR00T N1, Helix, HumanPlus, ASAP, OmniH2O — every recent humanoid foundation model that trained on more than 100 distinct scenes outperforms the same architecture on a curated 10-scene set. Generic 3D asset libraries cap the scene-diversity ceiling, and manual SimReady authoring caps it lower still.
Ego-vision and proprioception have to agree
Humanoid policies fuse first-person RGB / depth with proprioception. If a simulated object's mass and pose are even slightly off, the rendered scene and the kinematic readings disagree, and the policy learns the wrong correlation. Fixing the camera does not fix it; physics-accurate scene reconstruction does.
How Rigyd helps
AI-native infrastructure that automates the hard parts.
Calibration against the published humanoid kinematics
Outputs are tested against the reference humanoid URDF / MJCF used in Isaac Lab and MuJoCo Menagerie — Atlas (electric, 2024), Unitree H1 and G1, the NVIDIA Isaac GR00T reference humanoid, Tesla Optimus open-source replicas, and Digit (bipedal). Mass and friction calibration land inside each task's default domain-randomisation band, so policies trained on Rigyd assets transfer like those trained on hand-tuned assets.
Agentic scene-scale diversity
Generate thousands of distinct kitchen, office, warehouse, and household scenes from a single CAD or photo set. Each scene is per-object interactable — drawers open, chairs slide, kettles pour — because foundation-model training needs scenes a humanoid can actually manipulate, not single-mesh rooms.
Native support for GR00T, HumanoidBench, BEHAVIOR-1K, and MuJoCo MJX
Outputs drop into Isaac Lab's humanoid manipulation and locomotion task suites, GR00T N1 / N1.5 training pipelines, HumanoidBench (27 whole-body tasks, Microsoft Research, 2024), the BEHAVIOR-1K household benchmark, and MuJoCo MJX humanoid environments — with no per-stack re-authoring.
whole-body articulation supported across reference platforms
native pipelines for the leading foundation-model training stacks
distinct interactable scenes generatable per CAD or photo input set
Train humanoid foundation models on scenes that match the real world
Upload household, office, warehouse, or industrial scenes and get physically accurate SimReady assets and worlds in minutes.
Frequently asked questions
Which humanoid platforms is Rigyd calibrated against?
Rigyd outputs are tested against the reference humanoid models used in Isaac Lab and MuJoCo Menagerie — Atlas (Boston Dynamics, electric 2024 spec), Unitree H1 and G1, the NVIDIA Isaac GR00T reference humanoid, Tesla Optimus open-source kinematic replicas, and the bipedal Digit platform. For commercial humanoids whose kinematics are proprietary (Figure 02, 1X NEO, Apptronik Apollo, Sanctuary Phoenix, Fourier GR-1), Rigyd outputs are platform-agnostic SimReady environment assets that drop in once the platform's URDF / MJCF is loaded alongside. Per-platform calibration data (default friction, restitution bands matching each platform's preferred simulator settings) is available for Enterprise customers.
How do Rigyd assets work with NVIDIA Isaac GR00T?
GR00T N1 and N1.5 are trained on physics-accurate whole-body data, much of it synthetic. The bottleneck for any team replicating or extending GR00T-style work is scene diversity plus per-object physics calibration — exactly what Rigyd produces. Outputs are OpenUSD with USDPhysics schemas, the format Isaac Lab and the GR00T training stack consume natively. The 40% real-world performance lift reported in the GR00T N1 benchmark (NVIDIA, 2025) depends on physically accurate training data — that requirement is what Rigyd was built to satisfy at API rate.
Can Rigyd generate scene-scale environments (kitchens, offices, warehouses), not just objects?
Yes. Upload a 3D scan, BIM model, or set of reference photos of a room and Rigyd processes every distinct object inside — furniture, appliances, fixtures, decor — into individual SimReady assets with shared scene transforms. The output is a USD scene referencing per-object physics-enabled prims, ready for Isaac Lab whole-body tasks (BEHAVIOR-1K, HumanoidBench) or MuJoCo MJX humanoid environments. The per-object decomposition is essential: a humanoid policy needs to manipulate individual objects (open a drawer, lift a chair, pour from a kettle), not interact with a single room-scale mesh. Rigyd preserves that interactability automatically, which is otherwise the slowest part of preparing a home-scale environment by hand.
Does Rigyd support dexterous manipulation with multi-finger humanoid hands?
Yes. Asset collision geometry is generated at the contact resolution dexterous hands require — Inspire-Robots, SCHUNK SVH, Shadow Hand, and the multi-finger hands shipped on Figure 02, Apollo, NEO, and GR-1. Rigyd uses collision-aware decomposition (closer to CoACD than V-HACD) so contact-rich tasks — in-hand re-orientation, fingertip grasping, tool use — resolve correctly per finger. For benchmarks like ManiSkill, BEHAVIOR-1K dexterous tasks, and DexMimicGen / RoboCasa humanoid evaluation sets, Rigyd assets land in the same accuracy band as hand-tuned reference assets, with the per-asset effort dropping from engineer-hours to minutes.
What scene diversity actually matters for humanoid foundation-model training?
Three axes, ranked by impact: (1) per-object physics variation — mass, friction, restitution sampled across a calibrated band — generally the highest-leverage; (2) per-scene state variation — same room, different drawer-open / chair-position / item-placement states; (3) cross-scene topology variation — kitchen, office, warehouse, residential. Rigyd produces all three: physics-randomised SimReady assets, scene-state randomisation via USD scene variants, and topology variation by ingesting different CAD or photo inputs. HumanPlus, ASAP, OmniH2O, and the GR00T training narratives all converge on the same finding: scale of diversified scenes beats fidelity of a single hand-tuned scene.
Related reading
In-depth guides on robotics simulation, OpenUSD, and SimReady assets.
9 JUL 2026
Scaling Simulation Asset Libraries Beyond Curated Inventory
Curated SimReady libraries are a great starting point but a hard ceiling. Here is why fixed inventories limit robotics simulation at scale, and how on-demand asset generation closes the gap.
2 JUL 2026
Domain Randomization for Robotics Training: Asset Diversity at Scale
Domain randomization makes trained policies transfer to the real world, but it only works if your asset library is diverse enough. Here is how asset diversity drives randomization, and how to generate it at the scale training needs.
30 JUN 2026
Building a Scalable Embodied AI Asset Pipeline: From Raw Data to Simulation
A practical look at the stages of an embodied AI asset pipeline, from raw 3D data and reference inputs to physics-ready simulation assets, and what changes when you need to produce thousands of them instead of a handful.
Explore more
SimReady for Isaac Sim
OpenUSD assets validated for NVIDIA Isaac Sim
Models for MuJoCo
Physically accurate assets for robot learning
Assets for Gazebo
Simulation-ready models for ROS 2
Warehouse Simulation
50,000+ SKUs need physics properties
Humanoid Robots
Whole-body interaction training assets
Sim-to-Real Transfer
Close the gap with accurate physics data