Recursive Simulations · 2026-04-28

🔬 Recursive Simulations — 2026-04-28

🏭 Hannover Messe 2026: Industrial Digital Twins Cross the Line from Descriptive to Prescriptive
☁️ NVIDIA-Google Cloud Makes Physical AI Simulation Cloud-Native — Isaac Sim to Production Pipeline Closes
🚗 Unposed-to-3D (CVPR 2026): Autonomous Driving Simulation Learns Its Own Assets from Real-World Images
🔍 FunRec (CVPR 2026): Functional Scene Reconstruction Closes the Semantic Gap in Robotic Digital Twins
🏗️ From Visual Synthesis to Interactive Worlds: The Production-Ready 3D Asset Standard Is Converging
🚦 QNX Safety Stack Deploys on NVIDIA IGX Thor — But ISO/IEC 61508 Remains Structurally Blind to Learned Components

---

🏭 Hannover Messe 2026: Industrial Digital Twins Cross the Line from Descriptive to Prescriptive

Hannover Messe 2026, running April 20–24 in Germany, concentrated the most consequential shift in industrial simulation in a decade: digital twins are no longer design tools but operational ground truth. The signal is not any single deployment — it's the convergence of at least eight major industrial deployments all using simulation as the authority layer before any physical system goes live.

The clearest case: Humanoid's HMND01 wheeled humanoid, running NVIDIA Jetson Thor compute, developed entirely using Isaac Sim and Isaac Lab for simulation and reinforcement learning. Deployment to a Siemens electronics factory in Erlangen compressed hardware development from what typically takes two years down to seven months. The performance of the physical robot was effectively certified by simulation — not by field testing. That is authority inversion: the simulation is the source of truth, the physical world is the validation step, not the design environment.

SCHUNK's GROW automation cell extends this pattern to standardized form. NVIDIA Omniverse and Isaac simulation frameworks enable robot behavior to be simulated, trained, and validated before the cell goes live, with Wandelbots' NOVA platform providing continuous refinement post-deployment. ABB Genix integrates Omniverse with Microsoft Azure to let AI agents conduct root-cause analysis of asset performance through the digital twin rather than the physical asset — meaning operational decisions run through simulation infrastructure. Kongsberg Digital's Kognitwin extends this to energy infrastructure: test scenarios virtually, optimize before changes reach the physical world.

The operational question these deployments raise is epistemological: how do you audit a decision made by a system whose ground truth is the simulation? When Kognitwin recommends a valve configuration change and an operator implements it, what is the failure mode if the simulation's physics model is wrong? The entire cascade — training data, validation, commissioning, operational monitoring — runs through infrastructure that cannot currently be formally verified for the learned components it relies on.

The Hexagon Robotics AEON deployment at BMW Leipzig — one of the first humanoid deployments in a German production environment — makes this concrete. BMW is a safety-critical manufacturing environment. AEON's policy was trained in simulation. The question is not whether it works; the question is the governance architecture when it fails in ways the simulation did not anticipate.

What Hannover Messe makes visible is that simulation authority is now an organizational and governance problem, not just an engineering one. The industrial ecosystem has moved faster than the regulatory layer.

Sources:

---

☁️ NVIDIA-Google Cloud Makes Physical AI Simulation Cloud-Native — Isaac Sim to Production Pipeline Closes

At Google Cloud Next in Las Vegas (week of April 22), NVIDIA and Google Cloud announced the availability of NVIDIA Isaac Sim and Omniverse libraries on Google Cloud Marketplace — completing a pipeline that runs from CAD to simulation to cloud-scale robot training to physical deployment entirely without leaving managed cloud infrastructure. The announcement arrives alongside NVIDIA Vera Rubin A5X bare-metal instances on Google Cloud delivering 10x lower inference cost per token than the prior generation.

The architectural significance is distributional: until now, simulation-to-deployment pipelines required bespoke on-premises infrastructure to run Isaac Sim at scale alongside model training. Cloud availability lowers the capital barrier to entry for physical AI development from tens of millions in hardware to GPU-hours. Any organization with Google Cloud access can now iterate on physical AI in the same environment that runs production inference.

NVIDIA Cosmos Reason 2 NIM microservices are now deployable to Google Vertex AI and Google Kubernetes Engine, enabling vision-reasoning pipelines for robots and factory vision agents — trained on simulated scenarios — to run alongside production workloads. The Microsoft Azure Physical AI Toolchain, built on NVIDIA's Physical AI Data Factory Blueprint, extends the pattern to Azure.

The validation problem sharpens here. Schrödinger — a drug discovery company — is already using NVIDIA accelerated computing on Google Cloud to compress weeks-long molecular simulation into hours. That use case is relatively tractable: simulation correctness is measurable against known chemistry. For physical robot policy training, the simulation-to-cloud pipeline accelerates iteration cycles, but does not resolve the distribution shift problem: behaviors learned in cloud-hosted simulation environments must still transfer to physical hardware operating in uncontrolled environments. CrowdStrike's deployment of NeMo Data Designer and Automodel to generate synthetic cybersecurity training data follows a similar abstraction over replication logic — synthetic distributions designed to capture threat patterns rather than replicate production traffic directly.

The cloud-native physical AI pipeline completes a vertical integration that NVIDIA has been assembling since GTC 2026: chip (Blackwell/Rubin) → simulation (Isaac Sim) → model (Cosmos, Nemotron) → cloud (GCP/Azure) → edge (Jetson, IGX Thor) → physical deployment. The gap between each layer is now managed infrastructure rather than custom engineering work. The lock-in dynamics are structural: switching simulation environments mid-pipeline means retraining physical AI policies from scratch.

Sources:

---

🚗 Unposed-to-3D (CVPR 2026): Autonomous Driving Simulation Learns Its Own Assets from Real-World Images

The domain gap between synthetic training data and real-world deployment is the central validity problem in autonomous driving simulation. Unposed-to-3D, accepted to CVPR 2026 (Liu, Zou, Liu, Yu et al., April 21), addresses this by inverting the data generation question: instead of modeling synthetic vehicles and tuning them toward photorealism, the system learns to reconstruct simulation-ready 3D vehicles directly from unposed real-world images — photographs taken from arbitrary angles without known camera geometry, available at internet scale.

The "unposed" qualifier is technically significant. Standard 3D reconstruction requires multi-view sequences or depth sensors — structured capture conditions expensive to scale across the diversity of vehicle types, lighting conditions, and viewpoints autonomous driving simulation needs. Unposed-to-3D works from unstructured image collections at arbitrary angles. Output is geometrically accurate, physically meaningful, and importable into simulation environments like NVIDIA Isaac Sim and NVIDIA Omniverse. CVPR 2026 acceptance confirms reconstruction quality clears peer-reviewed standards for driving simulation research.

The distributional significance is where the analysis becomes interesting. Current autonomous driving simulation pipelines (CARLA, SUMO, NVIDIA Omniverse-based deployments) use artist-modeled or procedurally generated vehicle assets. These introduce systematic visual biases — idealized geometry, lighting, material approximations — that cause perception models to fail in specific and predictable ways on real imagery. Unposed-to-3D generates simulation assets from the same real-world image distribution that deployed models encounter: not merely higher fidelity but distribution-matched fidelity, where the simulation inherits real-world statistics rather than manually specified approximations.

The validation problem transforms accordingly. The question shifts from "does the simulation look real enough?" to "does the reconstruction introduce its own systematic biases?" Unposed reconstruction from internet-scale imagery inherits biases in what vehicles are photographically documented: rare commercial configurations, regional vehicle types, emergency vehicles in specific operating contexts, damaged vehicles. These underrepresented categories are precisely those where autonomous driving perception most needs reliability. The simulation pipeline inherits coverage gaps from internet photography rather than from artist modeling choices — a shift in failure mode source, not an elimination.

This is the bellwether for the NVIDIA Physical AI Data Factory Blueprint trajectory announced at Hannover Messe: treating simulation data generation as a production process with quality criteria. Unposed-to-3D provides a production ingredient that inherits data quality questions from its source distribution. The coverage problem doesn't disappear; it migrates from the simulation designer to the training image corpus.

Sources:

---

🔍 FunRec (CVPR 2026): Functional Scene Reconstruction Closes the Semantic Gap in Robotic Digital Twins

FunRec, accepted to CVPR 2026 and submitted to arXiv on April 26, 2026 (Delitzas, Zhang, Gavryushin, Di Mario, Guibas, Pollefeys, Engelmann, Barath et al.), addresses a foundational gap in digital twin methodology: existing reconstruction methods recover geometry and appearance, but not function. A robot trained on a geometric reconstruction of a kitchen knows where the drawer is but not that it can be opened, how it resists force, or what affordances it exposes for manipulation.

FunRec introduces a method for reconstructing functional 3D scenes from egocentric interaction videos — everyday people-using-things footage — and recovering not just object geometry but the functional interaction structure: what actions are possible, where affordances exist, how objects resist or yield to force. The reconstruction pipeline ingests video, identifies object-interaction events, and generates scene representations that encode functional properties alongside visual and geometric ones.

The practical implication for digital twins is direct: geometric twins tell simulation systems where things are; functional twins tell physical AI systems what can be done with them. For a robot navigating a factory floor or performing assembly operations, the difference is collision avoidance versus dexterous manipulation. The NVIDIA industrial digital twins stack at Hannover Messe — ABB Genix, Kongsberg Kognitwin, Siemens Digital Twin Composer — all currently operate at the geometric twin layer: they model asset positions and states, but not the functional affordances that tell a physical AI agent how to interact with those assets. CVPR 2026 acceptance signals the field is converging on functional reconstruction as the next required layer.

The validation question for FunRec is how functional fidelity degrades across object categories. Affordance reconstruction from video requires inferring physical properties — mass, compliance, friction — from visual information about human interaction. This inference is reliable for common objects with well-represented interaction patterns in training data, but systematically fails for novel objects, unusual configurations, or low-friction surfaces where human interaction patterns don't disambiguate material properties. A digital twin derived from FunRec-style reconstruction is only as functional as the interaction coverage in the source video.

The deeper structural issue: as functional reconstruction scales to industrial digital twins — the clear trajectory given Hannover Messe's deployment patterns — the authority question inverts. A functional twin that says a component can support 200kg of load is not just a visualization; it is a design constraint. If the functional twin's compliance model is wrong, the error is not a rendering artifact but a structural failure in production. The gap between functional representation and physical certification is the next friction point in simulation infrastructure deployment.

Sources:

---

🏗️ From Visual Synthesis to Interactive Worlds: The Production-Ready 3D Asset Standard Is Converging

Wu, Lou, Liu, Du, Guo et al.'s 41-page survey, "From Visual Synthesis to Interactive Worlds: Toward Production-Ready 3D Asset Generation" (April 26, 2026), maps a trajectory visible across this week's industrial and academic moves: 3D content generation is converging from isolated visual shapes toward structured assets deployable in real-time interactive environments. Three previously separate demand vectors are collapsing into one output specification: game asset pipelines, embodied AI training environments, and world model substrate generation.

The convergence is analytically important. Game pipelines optimize for visual fidelity at real-time frame rates. Embodied AI training environments need physical accuracy: collision geometry, mass properties, material interactions that survive contact. World model substrate generation needs semantic structure: what objects mean functionally, how they relate causally. These requirements have historically been served by separate workflows with separate quality criteria. Production-ready 3D asset generation is converging them into a single output specification that serves all three simultaneously.

What "production-ready" means is being standardized de facto by infrastructure providers. NVIDIA Isaac Sim defines what simulation-ready assets must provide (physics properties, collision meshes, joint definitions). OpenUSD defines the interchange format. The Hannover Messe 2026 deployment stack — Siemens Digital Twin Composer, ABB Genix, Wandelbots NOVA, Kongsberg Kognitwin — all use Omniverse + OpenUSD as the substrate for digital twin assets. This de facto standardization is happening faster than formal standards bodies (ISO, IEC, ANSI) can formalize it.

The lock-in dynamic this creates is structural. Assets generated to satisfy the NVIDIA/OpenUSD production-ready specification encode physics approximations chosen by that ecosystem. NVIDIA Cosmos world foundation models, trained on OpenUSD-standard assets, will develop physics intuitions that match OpenUSD's modeling choices — not necessarily ground-truth physics. This is not a defect: no physics engine models reality exactly. But it means the physics priors embedded in world models are determined by infrastructure provider choices, not by physics validation processes.

The survey also maps the unsolved scalability problem: maintaining physical fidelity when generating millions of assets for world model training. Current approaches achieve high fidelity for individual assets but degrade at scale. Hannover Messe's industrial digital twin deployments are gated by this — each factory-floor twin requires manually validated physics models for critical assets. Production-ready generation automation would accelerate deployment significantly, but validation methodology for automated physics fidelity at world-model training scale doesn't yet exist.

Sources:

---

🚦 QNX Safety Stack Deploys on NVIDIA IGX Thor — But ISO/IEC 61508 Remains Structurally Blind to Learned Components

At Hannover Messe 2026, QNX announced that QNX OS for Safety 8.0 is now integrated on NVIDIA IGX Thor and the NVIDIA Halos safety stack — the most serious attempt yet to bring formal safety architecture to AI-driven industrial robotics and medical systems. IGX Thor provides industrial-grade edge compute with functional safety guarantees. QNX provides a real-time operating system with safety certification. Halos provides system-level safety monitoring. Together they form a functional safety architecture for edge AI deployments.

The gap is structural and has not been addressed. ISO/IEC 61508, the foundational functional safety standard for industrial electronics, was designed for deterministic systems: logic that maps inputs to outputs via defined, auditable pathways. Safety integrity levels (SIL 1–4) are calculated based on probability of dangerous failure per hour, derived from fault tree analysis of deterministic logic. Simulation-trained neural network components — the actual decision-making layer in every physical AI system deployed at Hannover Messe — cannot be certified under 61508's current framework because their behavior under edge-case inputs cannot be formally enumerated. QNX certifies the OS layer; Halos monitors system-level behavior; but neither certifies what the learned model does in scenarios outside its training distribution.

ISO 26262 (automotive functional safety) has begun addressing AI components through its SOTIF (Safety Of The Intended Functionality) framework and ISO/PAS 8800 — AI safety in road vehicles — but these are partial and sector-specific. IEC 61511 (process industry) has no equivalent. The entire industrial simulation stack — from Isaac Sim training to Omniverse validation to IGX Thor deployment — is producing physical systems that operate at factory scale, in energy infrastructure, and in BMW production lines, under a safety governance framework designed for a prior technological paradigm.

What this means operationally: every humanoid robot at Hannover Messe running simulation-trained policies exists in a legal gray zone for safety-critical manufacturing deployment. QNX OS for Safety 8.0 + NVIDIA Halos is a genuine step toward safety architecture — it provides deterministic scheduling, fault isolation, and system-level monitoring that reduces risk substantially compared to running uncertified stacks. But it is not certification of the learned model, and the distinction matters enormously when failure cascades to physical injury or production line shutdown.

The bellwether case is AEON at BMW Leipzig: one of the first humanoids in a German production environment, per NVIDIA's announcement. BMW's internal safety engineers are likely treating this as a controlled experiment under rigorous human supervision. The question is what happens when commercial pressure accelerates deployment before new certification standards emerge — and whether IEC TC65 and ISO TC22's current AI working groups move fast enough to match the deployment velocity visible at Hannover Messe this week.

Sources:

---

Research Papers

Unposed-to-3D: Learning Simulation-Ready Vehicles from Real-World Images — Liu, Zou, Liu, Yu et al. (CVPR 2026, April 21) — Reconstructs simulation-ready 3D vehicle assets from unposed real-world images, closing the synthetic-to-real domain gap by generating training assets that inherit real-world image statistics rather than artist-modeled approximations.

FunRec: Reconstructing Functional 3D Scenes from Egocentric Interaction Videos — Delitzas, Zhang, Gavryushin, Guibas, Pollefeys, Engelmann, Barath et al. (CVPR 2026, April 26) — Reconstructs not just geometry but functional affordances and interaction structure from video, enabling digital twins that encode what agents can do with objects rather than just what objects look like.

From Visual Synthesis to Interactive Worlds: Toward Production-Ready 3D Asset Generation — Wu, Lou, Liu, Du, Guo et al. (April 26) — 41-page survey mapping the trajectory toward production-ready 3D asset generation convergent across game dev, embodied AI, and world model substrate demands; identifies the unsolved scalability problem for maintaining physics fidelity at world-model training scale.

SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds — Zhou, Liu, Jiang, Shen, Pang et al. (April 9–10) — Zero-shot transfer for deformable manipulation training using physics-aligned simulation; demonstrates that calibrated physics fidelity, not dataset size, is the binding constraint for synthetic-to-real transfer in non-rigid manipulation regimes.

World Simulation with Video Foundation Models for Physical AI (Cosmos-Predict2.5) — NVIDIA et al. (February 2026) — Introduces Cosmos-Predict2.5, trained on 200M curated video clips and refined with RL-based post-training, as a world foundation model for physical AI simulation; integrates with Cosmos-Reason1 for richer physics-grounded text control.

---

Implications

The convergence visible this week has a single through-line: simulation is becoming prescriptive infrastructure. Humanoid robots at BMW are commissioned in simulation and validated there before physical deployment. Autonomous driving perception models train on simulation assets derived from real imagery (Unposed-to-3D). Factory-floor digital twins (ABB Genix, Kongsberg Kognitwin) execute operational decisions through simulation pipelines before physical implementation. Functional reconstruction (FunRec) encodes affordances directly into the twin's representation layer. The production-ready 3D asset standard converging around OpenUSD is determining the physics priors embedded in world foundation models.

The authority inversion this implies is structurally significant. In conventional engineering, the physical world is the reference: simulations approximate it, designs are validated against it, systems are certified by operating in it. The infrastructure assembling around NVIDIA's Omniverse + Isaac + Cosmos stack inverts this order — the simulation is designed to be the reference, the physical world is the validation step, and the operational decision loop runs through simulation infrastructure. This inversion is now operative across industrial, automotive, and energy domains simultaneously.

A new analytical frame emerges from synthesizing Story 5 (production-ready 3D asset convergence) with Story 6 (certification gap): the physics priors embedded in world models are being determined by infrastructure provider choices, not by physics validation processes — and those priors are legally unauditable under current safety standards. The production-ready 3D asset specification converging around OpenUSD encodes specific physics approximations (collision models, material deformation, joint mechanics). NVIDIA Cosmos models trained on OpenUSD-standard assets learn physics intuitions derived from those approximations. ISO/IEC 61508 has no mechanism to evaluate whether those inherited physics priors are adequate for safety-critical use. The result is an unauditable physics authority chain: infrastructure provider → physics approximation → training distribution → world model priors → physical system behavior → deployed safety context. Each link is a technical choice; none is a certified physical reality. Automotive standards (ISO/PAS 8800) are beginning to address this for the road vehicle sector. Industrial equivalents don't exist.

The cloud-native pipeline closing (Isaac Sim on GCP, April 22) accelerates the timeline in a specific way: it decouples the barrier to entry from capital infrastructure, meaning the population of organizations deploying simulation-trained physical AI grows faster than regulatory apparatus can track. ISO/PAS 8800 took approximately 8 years from inception to publication. IEC TC65's AI working group has no published timeline for an industrial equivalent. The Hannover Messe deployments — BMW Leipzig humanoid, ABB energy infrastructure digital twins, Kongsberg oil and gas monitoring — are happening now. The bellwether event is not which simulation platform wins. It is whether the first fatal failure of a simulation-trained industrial system triggers regulatory response before standards bodies complete their deliberations.

---

HEURISTICS

`yaml heuristics: - id: simulation-authority-inversion domain: [industrial-AI, robotics, digital-twins, safety-governance] when: > Simulation-first development pipelines are deploying physical systems at scale. Digital twins are used as operational ground truth for robot commissioning, not as design visualization. Authority flows from simulation to physical, not the reverse. Industrial examples: Humanoid HMND01 (7-month dev via Isaac Sim), ABB Genix (root-cause via digital twin), Kongsberg Kognitwin (energy ops). prefer: > Analyze which layer holds decision authority: is simulation informing physical decisions (descriptive), or are physical decisions contingent on simulation approval (prescriptive)? Map authority flows explicitly. Flag where simulation outputs become binding constraints downstream — commissioning gates, operational triggers, maintenance decisions. Identify physics model scope: which physical regimes are covered, which are assumed to generalize. Out-of-distribution failure modes are invisible by definition until physical deployment surfaces them. over: > Treating simulation-to-real gap as a purely technical problem (domain randomization, higher fidelity rendering). The authority inversion is an organizational and governance problem, not just an accuracy problem. A simulation that is 95% accurate in training distribution produces policies that fail systematically on the 5% — and the failure cases are precisely those underrepresented in the training distribution. because: > NVIDIA Hannover Messe 2026: at least 8 industrial deployments (ABB, Siemens, SCHUNK, Kongsberg, Humanoid, Hexagon, Wandelbots, Invisible AI) use simulation as operational authority layer. Humanoid HMND01: 2 years → 7 months compression via simulation-first development. SIM1 (April 2026): zero-shot deformable manipulation training without real-world data. This is the production reality, not a research projection. breaks_when: > Physical environment differs systematically from simulation calibration domain (new materials, novel object configurations, environmental conditions outside training distribution). Failure modes are systematic not random — entire classes of scenarios fail together. Also breaks when simulation platform vendor changes, requiring full policy retraining. confidence: high source: report: "Recursive Simulations — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1

- id: simulation-certification-gap domain: [safety-governance, industrial-AI, robotics, regulatory] when: > Simulation-trained AI systems deploy in safety-critical environments: industrial robotics (ISO/IEC 61508), automotive (ISO 26262 / ISO/PAS 8800), medical devices (IEC 62304), process industry (IEC 61511). QNX OS for Safety 8.0 + NVIDIA Halos on IGX Thor: certifies OS layer and system monitoring, not learned model behavior. Humanoid robots at BMW Leipzig: first German production deployment of simulation-trained humanoid. BMW = safety-critical environment. prefer: > Distinguish three layers of safety architecture: (1) OS/scheduling certification (QNX, RTOS guarantees — achievable), (2) system-level monitoring (NVIDIA Halos — partial coverage), (3) learned model behavioral certification (ISO/IEC 61508 SIL-level — structurally impossible under current standards). Assess which layer a given deployment has certified. Identify the gap between what is certified and what makes safety-critical decisions. Treat ISO/PAS 8800 (AI in road vehicles, 2025) and IEC TC65 AI working group as regulatory leading indicators for industrial sector standards timeline. over: > Claiming simulation-trained systems are "certified" based on OS-layer or monitoring-layer certification alone. QNX certification ≠ learned model certification. NVIDIA Halos monitoring ≠ SIL-level behavioral guarantee. The gap between these is where liability and governance risk concentrates. because: > ISO/IEC 61508: SIL levels defined for deterministic logic via fault tree analysis — no pathway to certify neural network components. ISO 26262 SOTIF framework: partial AI guidance but sector-specific to automotive. IEC 61511 (process industry): no AI guidance. Every humanoid robot at Hannover Messe 2026 running sim-trained policies operates in legal gray zone for safety-critical deployment. Cloud-native physical AI pipeline (Isaac Sim on GCP, April 2026) accelerates deployment velocity faster than standards bodies can track. breaks_when: > New standards specifically designed for learned system components (ISO/PAS 8800 derivatives for industrial, medical) are adopted and enforce behavioral testing regimes for out-of-distribution scenarios. Also breaks if deployment regimes remain non-safety-critical (warehouse optimization vs. assembly near humans) — the gap matters more when failure consequences escalate. confidence: high source: report: "Recursive Simulations — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1

- id: physics-grounded-synthetic-data-threshold domain: [robotics, synthetic-data, embodied-AI, sim-to-real] when: > Training data collection for robotic manipulation is cost-prohibitive due to physical variability (deformable objects, novel materials, rare configurations). SIM1 (April 2026): deformable manipulation training via physics-aligned simulator, zero real-world data. ManiTwin (March 2026): 100K digital objects for manipulation training. PhysInOne (April 2026): physics reasoning from synthetic scenarios. FunRec (April 2026): functional affordance reconstruction from video. prefer: > Assess physics simulator calibration quality first: does simulation correctly model the specific physical regime (deformable soft matter, fluid-structure interaction, contact mechanics for materials outside engine defaults)? Generic simulators (Unity, Unreal, uncalibrated PhysX) fail systematically for non-rigid scenarios. Physics-aligned calibration (SIM1 approach) requires reference material data — identify whether calibration scope matches deployment scope. Synthetic data diversity increases coverage; physics fidelity determines transfer quality. These are independent variables. over: > Treating all synthetic data as equivalent regardless of physics fidelity. Zero-shot transfer results (SIM1) are regime-specific — deformable object results do not generalize to all robot manipulation. Each physical regime (granular, fluid, biological tissue, composite materials) requires separate calibration validation. because: > SIM1 (April 2026): zero-shot deformable manipulation training succeeds specifically because physics-aligned calibration matches material deformation dynamics. Generic simulators fail on deformable transfer due to contact model mismatches. Synthetic data exceeding real-world training data is a threshold phenomenon: reliable above physics fidelity floor, unreliable below it. No universal floor value — regime-dependent. breaks_when: > Physical environment introduces material combinations or interaction dynamics outside simulator calibration scope. Environmental factors (humidity, temperature, surface contamination) alter material properties in ways simulator calibration does not capture. Novel object categories with no reference calibration data available. confidence: medium source: report: "Recursive Simulations — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1 `