Introducing SimReadyGen — Agentic Simulation Generation for Physical AI

TL;DR

Lightwheel launched SimReadyGen, a text-prompt-to-SimReady-asset engine that generates OpenUSD assets carrying measured (not estimated) physics parameters for robot training. Every asset comes out of a two-stage stack: a Physics Measurement Factory that captures real-object contact, friction, and dynamics; and SimReady Foundry, which turns those measurements into ground-truth simulation parameters and also drives Lightwheel’s physics solver. The engine is built on top of NVIDIA Omniverse Content Agents so generated assets flow directly into Isaac Sim / Isaac Lab, and it sits inside Lightwheel’s broader “generate → evaluate (RoboFinals) → deploy (RoboStack) → learn” loop for Physical AI.

Key claims

• SimReadyGen generates structured, simulation-ready 3D assets from a text prompt, built on OpenUSD and integrated with NVIDIA Omniverse Libraries [§Introduction].
• Assets carry measured physics parameters (contact, friction, dynamics) sourced from Lightwheel’s Physics Measurement Factory, rather than estimated or LLM-guessed values [§Measured Physics].
• The same measurements are used both to parameterize generated assets and to develop Lightwheel’s physics solver, so simulated behavior is aligned with real object behavior end-to-end [§Measured Physics].
• Integration with NVIDIA Omniverse Content Agents automates the USD workflow — material assignment, physics property classification, texture generation, and content validation — so generated assets can move into Isaac Sim and Isaac Lab without manual post-processing [§Built for OpenUSD Workflows].
• SimReadyGen is the entry point of a closed continuous-learning loop: generated assets feed RoboFinals (evaluation platform), which feeds RoboStack (deployment), and real-world performance data returns to refine the next round of simulation [§Lightwheel’s Continuous Learning System].

Method

The announcement describes a two-layer stack rather than a single model. At the bottom, a Physics Measurement Factory physically measures real objects (contact behavior, friction coefficients, dynamics properties). Those measurements feed SimReady Foundry, which converts them into ground-truth USD physics parameters and also drives the development of Lightwheel’s own physics solver — closing the loop between measurement and simulation semantics. At the top, SimReadyGen is the generation-time interface: a text prompt is turned into an Omniverse-Content-Agents-driven pipeline that assembles the USD asset (geometry, materials, textures, physics classification, validation) and draws physics parameters from the Foundry’s measured library. Assets emerge as OpenUSD files ready to load into Isaac Sim / Isaac Lab. The post does not disclose model architectures, training data, or scale/latency numbers.

Results

The announcement is product-launch framing rather than a benchmark report — no quantitative sim-to-real transfer numbers, no dataset scale, no policy-training results are given. The evaluative claim reduces to “measured physics, not estimates” and integration with Omniverse / Isaac. The demonstrated artifact is the SimReadyGen product itself (available at lightwheel.ai/simreadygen/generate/asset), plus the promise that generated assets participate in the broader RoboFinals evaluation platform.

Why it’s interesting

SimReadyGen stakes out the “measured physics as generation-time constraint” position in the crowded synthetic-data-for-robotics space: where SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation (SimFoundry) generates affordance-preserving digital cousins of video-reconstructed twins and SAGE: Scalable Agentic 3D Scene Generation for Embodied AI (SAGE) generates task-conditioned scenes from text with LLM agents, Lightwheel argues the bottleneck is not scene composition but physical accuracy of the individual asset — and offers a measurement pipeline as the fix. It’s also a natural counterpart to The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward (Genesis World), which supplies the multi-physics simulator; SimReadyGen supplies the physically-parameterized content that fills those simulators. The Omniverse / Isaac integration puts it in the same delivery lane as NVIDIA’s NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI Cosmos 3 SDG stack, but positioned as an asset generator rather than a scene / video generator.

See also

• Synthetic Training Data — SimReadyGen is another entry in the “measured / physical simulation as training substrate” flavor.
• SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation — the video-to-digital-cousin recipe; SimReadyGen is the text-to-asset counterpart.
• SAGE: Scalable Agentic 3D Scene Generation for Embodied AI — text-to-scene agentic pipeline for embodied-AI simulation; complementary layer above SimReadyGen’s asset-level output.
• The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward — the multi-physics engine side of the same stack.
• NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI — NVIDIA’s competing / adjacent physical-AI synthetic-data engine, but at the video / trajectory layer rather than the asset layer.