Recursive Simulations · 2026-04-25

🔄 Recursive Simulations — 2026-04-25

🏗️ NVIDIA Omniverse Disaggregates: Simulation Infrastructure Becomes a Library Stack
🧵 FLASH and SIM1: Physics-Native Synthesis Defeats the Deformable Data Problem
⚛️ AI Surrogates Enter Nuclear Design: PhysicsNeMo Achieves R²=0.97 on Neutron Flux
🧪 ALCHEMI Toolkit: Atomistic Simulation as GPU Microservice Infrastructure
🤖 MCP Servers Install AI Agents as Simulation Operators
🚗 Unposed-to-3D Closes the Real-World-to-Simulation Asset Pipeline

---

🏗️ NVIDIA Omniverse Disaggregates: Simulation Infrastructure Becomes a Library Stack

The monolithic simulation platform is ending. On April 8, NVIDIA announced that its Omniverse capabilities—RTX rendering, PhysX simulation, and USD storage pipelines—are now available as standalone, headless-first C APIs: ovrtx, ovphysx, and ovstorage. Industrial teams can embed photorealistic rendering and physics directly into existing applications without adopting the full Omniverse Kit framework or triggering architectural rewrites.

The design decision signals what NVIDIA calls "physical AI"—AI systems trained and validated in physically grounded simulated environments—as a production workload rather than a research primitive. Six major industrial software companies are in early adoption or pilot integration: ABB Robotics is embedding the libraries into RobotStudio for offline programming and commissioning of industrial robots; PTC is connecting Onshape directly to Isaac Sim for cloud-native robot design and testing; Siemens is building industrial digital twins at scale via the DSX blueprint.

The breakup of the monolith is also visible in NVIDIA's own engineering decisions. Isaac Lab 3.0 Beta has transitioned from the Kit framework to a multi-backend architecture: teams can choose between ovphysx (wrapping PhysX) or a Kit-less Newton backend powered by MuJoCo-Warp for physics; OVRTX, Isaac RTX, and lightweight visualizers including Rerun and Viser are pluggable rendering alternatives. This breaks the assumption that robotics simulation means "running Omniverse"—it means choosing appropriate physics and rendering primitives for specific workloads.

The practical motivation is architectural. Isaac Lab previously required the Kit runtime loop, which introduced framework-induced latency and prevented explicit execution control. Standalone libraries enable deterministic stepping, decoupled update frequencies (high-frequency IMUs and lower-frequency vision systems operating at native rates within one environment), and tensorized data exchange for direct GPU buffer access—positions and velocities as PyTorch tensors without host copies. These are not quality-of-life improvements: they are preconditions for deploying simulation inside CI/CD pipelines, inside edge compute stacks, inside real-time control loops.

What this structural shift implies: simulation infrastructure is disaggregating the same way cloud compute disaggregated into primitives (compute, storage, networking) in the 2010s. The Omniverse platform may remain the composition layer for new applications, but the constituent capabilities—physics, rendering, spatial data—are becoming independently addressable. Cadence, Adobe, and Synopsys are each embedding specific capabilities into their existing stacks without full-platform adoption. The unit of analysis is no longer "are you using Omniverse?" but "which Omniverse primitive handles which layer of your simulation stack?"—a question that implies vendors, not NVIDIA, now architect the physical AI substrate.

Sources:

---

🧵 FLASH and SIM1: Physics-Native Synthesis Defeats the Deformable Data Problem

The canonical bottleneck in embodied robot learning isn't hardware or model capacity—it's data. Deformable object manipulation (cloth, food, biological tissue, flexible packaging) sits at the hardest end: shape, contact, and topology co-evolve in ways that real-world teleoperation cannot easily sample. Two papers from April 2026 directly attack this from the physics-simulation side.

SIM1 (Zhou et al., April 10, 2026) frames physics-aligned simulation as a "zero-shot data scaler"—a system that generates training data for deformable manipulation without any real-world demonstrations. The argument is that rigid object manipulation has largely solved its data problem through sim-to-real transfer, but deformable worlds remained intractable because prior simulators lacked the contact fidelity required for meaningful transfer. SIM1 addresses this through physics alignment: the simulator preserves the essential physical invariants (contact geometry, friction models, compliance) that determine whether policies transfer. The result is a zero-shot data pipeline—simulation data that transfers to real hardware without requiring target domain demonstrations.

FLASH (Luo et al., April 19, 2026) takes a different angle on the same problem: compression of the training loop. Using GPU-accelerated simulation built on Isaac Sim, FLASH demonstrates that high-fidelity deformable manipulation policies can be learned in minutes rather than the days-scale training cycles that characterize prior work. The key innovation is GPU parallelism applied to contact-rich, deformable object dynamics—a regime previously intractable for parallelization because of sequential contact resolution steps. FLASH's results suggest that the rate-limiting factor in deformable robot learning was compute substrate, not policy architecture.

The convergence pattern here is significant. Both papers treat physics simulation not as a research approximation but as a data source that must pass transfer fidelity tests. The standard question has shifted from "how realistic does this look?" to "do policies trained here transfer without fine-tuning?" This is precisely the transition from simulation as description to simulation as prescription—the simulator's physical model is the authority certifying training data quality.

The implications for safety-critical industrial automation are direct. Deformable handling is pervasive in pharmaceutical packaging, food processing, and surgical robotics—contexts where real-world data collection is expensive, slow, or legally restricted. If physics-aligned simulation can generate validated training distributions at GPU-scale speed, it eliminates one of the primary arguments for hardware-teleoperation data collection. The question that follows is validation methodology: when a deployed system fails at a deformable task, how do you determine whether the failure was physics-model error, distribution shift, or policy error? The deformable data pipeline is production-ready before the audit infrastructure for it exists.

Sources:

---

⚛️ AI Surrogates Enter Nuclear Design: PhysicsNeMo Achieves R²=0.97 on Neutron Flux

The highest-stakes claim in simulation epistemology is not "this looks realistic" but "this is safe." Nuclear reactor design has always required simulation—physical experiments are dangerous and impractical at scale—but the bottleneck is computational cost: high-fidelity Monte Carlo neutron transport simulations that resolve individual pin-cell flux distributions are too expensive for iterative design exploration. NVIDIA's April 17 workflow using PhysicsNeMo demonstrates that Fourier Neural Operators (FNOs) can replace expensive transport solves with near real-time surrogate predictions—achieving R²=0.97 versus R²=0.80 for the scalar regression baseline.

The technical argument is instructive. A typical reactor core contains ~50,000 fuel pins; full-core simulation at explicit pin-cell resolution is computationally impractical. Standard approaches condense fine-scale transport physics into homogenized cross-sections for coarser core simulators. The PhysicsNeMo workflow trains an FNO to jointly predict the spatial neutron flux field and the absorption cross-section field, then computes homogenized cross-sections from those fields—mimicking the physics-based aggregation that expensive Monte Carlo methods perform. The scalar baseline collapses the spatial information that determines flux weighting: geometrically similar pin cells can have significantly different flux distributions due to self-shielding. The FNO, operating on field representations rather than compressed scalars, preserves this spatial structure and achieves substantially better generalizability.

The epistemological shift is subtle but consequential. The FNO surrogate does not model physics—it learns to predict the outputs of a physics model at far lower cost. The physics model (Monte Carlo transport) remains the authority; the surrogate is a fast approximation. But when the surrogate is accurate enough, it becomes the de facto standard in design decisions: engineers explore the design space using the surrogate, and expensive Monte Carlo runs are reserved for validation checkpoints. The surrogate's predictions prescribe which regions of design space are worth exploring.

PhysicsNeMo integrates with CUDA-X libraries, the PhysicsNeMo Curator for training data pipelines, and NVIDIA Omniverse for digital twin deployment. This full-stack integration means the nuclear design workflow is not a research prototype—it is a reference architecture for bringing surrogate-based design into certification-adjacent workflows. The critical unresolved question is regulatory: current nuclear frameworks (10 CFR Part 50) have no established path for certifying designs validated primarily through AI surrogate models. The gap between computational capability and regulatory authority is widening at exactly the moment surrogates are becoming production-grade. ISO/IEC 61508 and its nuclear-sector derivatives require traceable, deterministic relationships between design parameters and safety outcomes—a Fourier Neural Operator's error distribution, however tight, does not satisfy that requirement.

Sources:

---

🧪 ALCHEMI Toolkit: Atomistic Simulation as GPU Microservice Infrastructure

Machine learning interatomic potentials (MLIPs) represent the clearest case of simulation abstraction over replication: instead of modeling every quantum interaction from first principles (DFT, ab initio), MLIPs learn the energy surface from sparse quantum calculations and generalize across chemical configurations at near-classical computational cost. DFT methods are limited to hundreds of atoms; classical force fields scale to millions but lose chemical accuracy; MLIPs sit at the intersection, providing quantum accuracy at GPU-scale throughput. NVIDIA's ALCHEMI Toolkit, released April 14, 2026, addresses the bottleneck that replaced the physics bottleneck: the software infrastructure surrounding MLIPs remains CPU-centric legacy code, throttling simulations that are computationally capable of GPU-native execution.

ALCHEMI Toolkit is a collection of GPU-accelerated simulation building blocks—neighbor list construction, DFT-D3 dispersion corrections, PME electrostatics for periodic Coulomb interactions, batched dynamics kernels—assembled as a PyTorch-native framework managing data flow between domain-specific GPU kernels and deep learning models. The design is explicitly composable: researchers build custom simulation workflows by combining primitives rather than adapting monolithic codebases. This is the atomistic-simulation version of what Omniverse's library decomposition does for physical AI: primitives replace platforms.

Three industry integrations define the current adoption curve. Orbital has integrated ALCHEMI Toolkit into OrbMolv2—achieving approximately 1.7x acceleration for large systems and 33x for batched smaller systems via TorchSim integration. Matlantis, combining universal MLIPs with cloud computing for materials discovery, has achieved up to 10x speedup via Warp-optimized neighbor list construction; full Toolkit integration could enable high-throughput parallel relaxation of millions of molecular configurations simultaneously. Materials Graph Library (MatGL) is integrating GPU-native kernels to accelerate graph-based MLIP workflows for materials property predictions at scale.

The pattern across all three integrations is identical: established frameworks with proven physics but CPU-bound infrastructure are grafting GPU-native kernels onto their data pipelines without replacing the underlying physics models. Simulation authority remains in the quantum mechanical models; computational substrate accelerates. This is "augmented simulation infrastructure"—the same physics, delivered faster.

The stakes for materials discovery scale directly with throughput. The bottleneck in battery electrolyte design, catalyst discovery, and drug-molecule property prediction is not accuracy—it is the number of configurations that can be explored per compute-hour. ALCHEMI Toolkit's throughput improvements transform design space exploration from thousands-of-structures to millions-of-structures, qualitatively changing what is discoverable. Whether this leads to better materials or faster exploration of a space already well-characterized by cheaper methods is an open empirical question—one that MLIP adoption in industrial pipelines will answer over the next 24 months. The validation challenge is symmetric: just as deformable robot data pipelines need audit infrastructure, high-throughput materials simulation needs screening protocols that distinguish genuine novel candidates from artifacts of the generalization regime.

Sources:

---

🤖 MCP Servers Install AI Agents as Simulation Operators

The authority structure of physics simulation has always been specialist-mediated: simulation engineers write scripts, compose scenes, configure physics parameters, interpret results. The workflow requires deep domain expertise—not primarily in physics but in the software stacks (USD scene graphs, PhysX parameter spaces, sensor model configurations) that mediate between intent and simulation output. NVIDIA's April 8 Omniverse post introduced a structural change less discussed than the library modularization: Omniverse capabilities are now exposed via Model Context Protocol (MCP) servers, making simulation operations callable by LLM-based agents.

The implementation involves three MCP server types: USD Code MCP (browsing APIs, generating scene code, manipulating layer hierarchies), Kit MCP (interacting with the Omniverse Kit framework), and OmniUI MCP (UI element manipulation). Each server describes simulation operations—load USD scene, edit prim properties, step physics simulation, retrieve sensor outputs—in machine-readable schema. Tools including Claude and Cursor can parse these schemas and issue simulation commands from high-level natural language instructions without specialist mediation.

The structural consequence is an inversion in who can operate simulation infrastructure. Previously, deploying a robot training environment required simulation-specialist labor: configure physics parameters, debug contact models, validate that the simulation matched the physical system being modeled. With MCP servers, a developer with no simulation expertise can instruct an LLM agent to "create a manipulation scene with a deformable cloth on a table, apply random wind forces, and collect 10,000 training episodes." The agent issues RPC calls to Omniverse; the simulation engineer's tacit knowledge about which physical parameters are load-bearing (friction coefficients, contact tolerances) versus decorative (lighting, texture fidelity) is now embedded in the LLM's training data.

This creates an underexamined validation problem. LLM agents orchestrating simulation through MCP servers do not discriminate between physics-critical parameters and cosmetic ones by default—they generate plausible-seeming configurations that may systematically mis-specify the physics properties that determine transfer fidelity. The NemoClaw integration mentioned alongside MCP servers suggests NVIDIA is aware of this constraint: agentic simulation workflows need policy-based guardrails, not just API access.

The decade-scale implication is that simulation expertise becomes embedded in LLM orchestration layers—which means the quality of physical AI training pipelines will depend on the quality of the models orchestrating them. When deployed systems fail at transfer, the debugging chain now includes "what did the agent specify when it configured the simulation?" This is a new kind of epistemic accountability problem: the simulation's physical fidelity is downstream of a language model's physics intuitions. Authority inversion here is not a feature—it is a risk surface that scales with the adoption of agentic simulation workflows.

Sources:

---

🚗 Unposed-to-3D Closes the Real-World-to-Simulation Asset Pipeline

The production physics simulation pipeline has a persistent upstream bottleneck: getting real-world objects into simulation-ready 3D representations. Traditional approaches require controlled capture rigs, multi-camera setups, and manual rigging—processes that work for curated asset libraries but break down when applied to the diversity of objects that appear in industrial or autonomous driving environments. Unposed-to-3D (Liu et al., April 21, 2026) directly attacks this bottleneck by learning to generate simulation-ready 3D vehicle assets from unposed, casually-captured real-world images—eliminating the domain gap that plagues prior synthetic methods.

The key insight is in what "simulation-ready" means. Prior 3D vehicle generation methods optimized for visual fidelity (photorealistic rendering) but generated assets that failed physics simulation because their geometry was not physically consistent—surfaces had gaps, interior volumes were not watertight, collision meshes were decorative rather than functional. Unposed-to-3D generates assets that pass simulation integrity checks: physically consistent geometry, appropriate contact surfaces, and consistent appearance under diverse sensor modalities (camera, LiDAR, radar). The domain gap between "looks realistic" and "behaves realistically under physics" is the precise gap the paper addresses.

This connects to a convergent pattern across April 2026 simulation developments: the distinction between real-world data and synthetic data is collapsing at the asset level. Cosmos Transfer 2.5 converts Omniverse simulation outputs into photorealistic video via ControlNet-based style transfer, bridging the visual gap from simulation to reality. Unposed-to-3D closes the gap in the other direction—converting real-world imagery into physics-simulation-ready geometry. Together they constitute a bidirectional bridge: the same content can exist in both regimes, converted as needed.

The implications for autonomous vehicle and industrial robotics training pipelines are significant. The current production workflow separates real-world data collection (expensive, slow, geographically constrained) from synthetic data generation (fast, scalable, controllable). The emerging architecture treats real-world captures as raw inputs to asset-generation pipelines, which produce physics-simulation-ready assets, which feed synthetic data generators producing diverse training distributions. The "real world" is no longer the ground truth—it is the upstream input to a simulation-mediated training pipeline.

What breaks when this fails is the question receiving insufficient attention. If the asset generation model systematically mis-represents vehicle contact geometry for low-frequency vehicle classes (unusual body types, modified vehicles), the trained policy's failures will be invisible until real-world deployment. The simulation pipeline was the authority on what the training distribution looked like—and it was wrong in ways the training data could not reveal. This is the simulation-prescriptive problem made structural: the asset generation model's biases propagate into every downstream policy trained on the synthetic data it generates, invisibly, at scale.

Sources:

---

Research Papers

FLASH: Fast Learning via GPU-Accelerated Simulation for High-Fidelity Deformable Manipulation in Minutes — Luo et al. (April 19, 2026) — GPU-parallel deformable physics in Isaac Sim reduces policy learning from days to minutes; demonstrates that contact-rich deformable manipulation was compute-substrate-limited rather than architecture-limited.

SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds — Zhou et al. (April 10, 2026) — Frames physics-aligned simulation as a zero-shot data pipeline for deformable manipulation, addressing the contact-fidelity gap that has blocked sim-to-real transfer in this regime.

Unposed-to-3D: Learning Simulation-Ready Vehicles from Real-World Images — Liu et al. (April 21, 2026) — Generates physically-consistent 3D vehicle assets from unposed real-world images; explicitly targets the domain gap between visual fidelity and physics-simulation integrity.

Telecom World Models: Unifying Digital Twins, Foundation Models, and Predictive Planning for 6G — Zou et al. (April 8, 2026) — Proposes a unified telecom world model architecture combining physics-based digital twins and foundation model prediction; demonstrates that simulation-prescriptive authority is extending from industrial to wireless infrastructure.

---

Implications

The week's simulation developments share a structural logic: disaggregation of the simulation stack is creating insertion points for AI authority at every layer, while the boundary between real-world data and synthetic data dissolves from both directions simultaneously.

The cross-thread synthesis that matters: simulation authority is moving up the stack while physics expertise migrates down into infrastructure. When PhysicsNeMo surrogates achieve R²=0.97 for nuclear reactor design, domain experts are no longer running simulations—they are validating surrogate predictions. When MCP servers allow LLM agents to orchestrate Omniverse physics without specialist mediation, simulation engineers' tacit knowledge is embedded in the LLM's training data. When Cosmos Transfer converts simulation outputs to photorealistic video and Unposed-to-3D converts real-world images to simulation-ready assets, the real/synthetic distinction collapses into a production pipeline detail. These are not independent developments—they are a coordinated authority transition in which the physics model is infrastructure and the LLM orchestration layer is the new operator.

The regulatory gap is the decade-scale problem none of these developments address. ISO/IEC 61508, the foundational standard for functional safety in industrial systems, was designed for deterministic, traceable simulation pipelines. A surrogate model trained on Fourier Neural Operators that achieves R²=0.97 on held-out test sets is not auditable under 61508's framework—the standard requires traceable, deterministic relationships between design parameters and safety outcomes. The same gap exists in aviation (DO-178C), automotive (ISO 26262), and nuclear regulation (10 CFR Part 50): certification frameworks require that every validation step be auditable, reproducible, and deterministically connected to the physical system being certified. None of the current simulation infrastructure—not PhysicsNeMo surrogates, not Cosmos world foundation models, not MCP-orchestrated physics environments—is certifiable under these standards.

This creates a specific lock-in dynamic. ABB, Siemens, PTC, Cadence, and Synopsys are integrating NVIDIA's simulation primitives into production industrial workflows now, before the regulatory frameworks governing those industries have addressed how to certify AI-augmented simulation pipelines. Each integration deepens the dependency on learned-model components. By the time standards bodies produce updated guidelines for AI-augmented simulation in safety-critical applications—a process that typically takes 5-7 years—the simulation stacks will be entrenched and compliance retrofitting will be prohibitive. The trajectory is simulation authority accreting faster than simulation accountability, across every domain simultaneously. The bellwether will be the first major industrial incident attributable to a policy trained on synthetic data generated by an AI-orchestrated simulation pipeline—that event will determine whether certification frameworks adapt or whether the industry absorbs the failure mode as acceptable operational risk.

---

HEURISTICS

`yaml heuristics: - id: simulation-prescriptive-authority domain: [simulation, safety-critical, certification, nuclear, robotics, automotive] when: > AI surrogate models replace expensive physics solvers in design workflows. Practitioners cite R² scores and error metrics as proxies for physical validity. FNOs, neural operators, or learned models generate outputs that directly feed downstream design decisions. Relevant domains: nuclear, structural, automotive, aerospace, pharmaceutical manufacturing. prefer: > Maintain hierarchical authority structure: learned surrogate as computational accelerant, first-principles solver as validation oracle. Schedule periodic Monte Carlo or FEM validation checkpoints at design decision nodes rather than only at end of design exploration. Instrument surrogate predictions with uncertainty quantification—FNOs can produce calibrated confidence intervals via MC dropout or ensemble methods. Track design-space regions where surrogate uncertainty exceeds domain-specific tolerance thresholds and route those regions back to expensive physics solvers. over: > Treating surrogate accuracy on held-out test sets as sufficient validation for safety-critical design decisions. R²=0.97 on a test distribution does not bound error on out-of-distribution geometries or operating regimes not present in the training set. Surrogate confidence is not physical validity. because: > PhysicsNeMo FNO achieves R²=0.97 vs R²=0.80 scalar baseline on nuclear pin-cell neutron flux prediction (NVIDIA, Apr 17 2026). Test distribution is Latin Hypercube sampled from bounded parameter space—not a coverage guarantee for novel enrichment geometries or accident scenarios. NRC 10 CFR Part 50 Appendix K requires traceable deterministic validation chains: a surrogate model's error distribution is not currently accepted as a certification artifact. ISO/IEC 61508 SIL-3 and SIL-4 requirements specify traceable, deterministic relationships between parameters and outcomes. breaks_when: > Design space is exhaustively characterized and surrogate extrapolation is never required. Independent first-principles validation is performed at each design decision node. Certification frameworks are updated to accept probabilistic surrogate validation with specified uncertainty budgets. confidence: high source: report: "Recursive Simulations — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1

- id: simulation-modularization-lock-in domain: [simulation, infrastructure, industrial, vendor-dependency, certification] when: > Simulation platforms decompose into composable library primitives. Industrial partners embed individual capabilities—physics, rendering, storage—into existing stacks independently. Adoption decisions made on per-layer capability fit rather than full-platform migration. Domain: automotive, manufacturing, aerospace, robotics, semiconductor design. prefer: > Map vendor dependency depth at each layer: physics kernel (ovphysx/PhysX), rendering (ovrtx/RTX), data format (OpenUSD), orchestration (MCP/Kit). Identify substitutability of each layer independently. Prioritize open standards (OpenUSD) over proprietary APIs at layers where certification or auditability requirements apply. Evaluate switching cost as sum of layer-by-layer migration complexity rather than total platform complexity. Classify layers as safety-certification-relevant vs. productivity-relevant before adoption—safety-relevant layers require vendor audit rights. over: > Evaluating simulation vendor lock-in as monolithic. Platform disaggregation enables selective, invisible lock-in at individual high-value layers. The lowest-substitutability layer (typically physics kernel or data format) becomes the actual lock-in surface regardless of what other layers are open. because: > ABB, Siemens, PTC, Adobe, Cadence, Synopsys all integrating Omniverse primitives into existing stacks at GTC 2026 (Apr 8 2026). Each integration embeds a specific Omniverse layer while preserving existing architecture. Isaac Lab 3.0 Beta demonstrates multi-backend physics (ovphysx vs MuJoCo-Warp)—proving physics layer is substitutable while rendering and OpenUSD data format converge on NVIDIA defaults across all six adopters. breaks_when: > Open alternatives emerge for RTX-equivalent photorealistic sensor simulation with independent certification history. OpenUSD ecosystem includes non-NVIDIA physics backends with equivalent performance. Regulatory requirements mandate specific physics solvers with independent certification. confidence: medium source: report: "Recursive Simulations — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1

- id: synthetic-real-boundary-collapse domain: [simulation, synthetic-data, computer-vision, robotics, autonomous-vehicles] when: > Asset generation models produce simulation-ready geometry from real-world captures. World foundation models convert simulation outputs to photorealistic video. Bidirectional real-to-sim and sim-to-real bridges coexist in the same production stack. Training pipelines treat real-world captures as upstream inputs rather than ground truth. Domain: autonomous vehicles, industrial robotics, surgical robotics. prefer: > Audit training distribution for systematic under-representation before pipeline deployment. Trace which real-world inputs seeded which asset classes in the simulation pipeline. Maintain out-of-distribution test sets sampled from real-world distributions not present in seed captures. Instrument deployed models with distribution-shift detectors calibrated to flag inputs unlike anything in the simulation training distribution. Separate visual fidelity validation (does this look real?) from physics fidelity validation (does policy transfer?). over: > Treating photorealism (visual similarity between sim and real) as equivalent to distribution coverage (physical diversity of scenarios encountered in deployment). High visual fidelity from Cosmos Transfer does not imply rare-event physics coverage. Asset generation bias propagates invisibly into every downstream policy. because: > Unposed-to-3D generates simulation-ready vehicle assets from unposed real-world images (Apr 21 2026). Cosmos Transfer 2.5 converts Omniverse output to photorealistic video via ControlNet (updated Mar 13 2026). When both bridges exist, training distribution is determined by the seed images used to generate assets—asset class frequency mirrors seed image availability, not deployment encounter rates. Low-frequency vehicle classes (unusual body types, modified vehicles) will be systematically under-represented relative to their operational hazard. breaks_when: > Asset generation is explicitly conditioned on target deployment distribution with rare-class oversampling. Out-of-distribution sampling is built into the generation pipeline. Rare-event coverage is independently verified against deployment telemetry before production release. confidence: high source: report: "Recursive Simulations — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1 `