The State of Play

Simulation-based robot training has reached production maturity within scoped domains, yet the fundamental reality gap that has constrained the field since inception remains unresolved and unavoidable. ICRA 2026 marked a consolidation milestone: NVIDIA's COMPASS framework achieved 80% real-world navigation success from simulation-only training; Grasp-MPC demonstrated 75% success on real robots trained on 2 million synthetic trajectories; SPARR assembly workflows showed 38% improvement in success rate and 30% cycle-time reduction. These results position simulation-based training as an engineering decision for teams building physical AI, not a research aspiration. Yet even as deployment accelerates, critical assessments quantify the remaining gap: 80% success is insufficient for safety-critical applications where a 20% error rate proves catastrophic; contact dynamics, friction, and sensor noise remain notoriously difficult to model faithfully; the gap between 80% and production-grade 99% reliability is often as large as the gap from zero to 80%. The field consensus is clear: simulation-based training is a high-value tool for perception, navigation, and scoped manipulation, but it requires careful task selection, acceptance of persistent reality gaps, and typically benefits from hybrid real-world fine-tuning or continuous correction mechanisms rather than zero-shot transfer.

ICRA 2026 demonstrated production-ready sim-to-real across multiple domains via GPU-accelerated training pipelines. NVIDIA's COMPASS framework trained entirely in Isaac Lab simulation achieved 80% real-world navigation success across 20 trials with zero physical data; Grasp-MPC reached 75% success on real robots from 2 million synthetic trajectories across 8,000 objects; SPARR assembly workflows showed 38.4% success-rate improvement and 29.7% cycle-time reduction on unseen NIST tasks. These results position simulation-based training as a production decision for rapid development iteration. ABB's RobotStudio HyperReality with Omniverse, FANUC-NVIDIA integration of Isaac Sim and ROBOGUIDE, and ecosystem adoption by Agility, Boston Dynamics, and Figure AI confirm vendor consolidation around GPU-accelerated simulation. Parallel methodologies have matured: HyperSim demonstrates holistic bridging via synthetic data generation, adversarial trajectory generation, and sim-and-real co-training (80–95% success, 35% robustness improvement under perturbation across 400 real-world executions); hybrid co-training and real-to-sim-to-real approaches reduce dependence on pure simulation; domain randomisation and learned dynamics models continue to narrow the gap in perception and navigation tasks.

Yet critical assessments temper optimism. An 80% success rate proves insufficient for safety-critical deployments; the "long tail" of unexpected events—handling slightly broken door mechanisms, adapting to production anomalies, maintaining repeatability under endurance—remains unresolved. Contact dynamics, friction, and sensor modelling remain notoriously difficult, with the gap between high-fidelity simulators and real hardware persisting at 24–30% performance degradation even on controlled tasks. Practitioners consistently note that closing the reality gap remains "one of the most pressing challenges in robotics" despite six years of methodological investment. The field consensus consolidates around scoped, task-specific adoption: simulation-based training excels for perception and navigation, shows strong capability in constrained manipulation, and remains constrained by persistent gaps in contact-rich, dynamic, and unstructured tasks. Deployment requires careful gap characterization, acceptance of non-zero failure rates, and typically benefits from hybrid real-world fine-tuning or continuous correction mechanisms rather than zero-shot transfer.

• 2018: Domain randomization emerges as a key technique for sim-to-real transfer; ICRA 2018 landmark paper from OpenAI/UC Berkeley demonstrates successful transfer on robotic manipulation (object pushing). Georgia Tech validates transfer on biped locomotion. NVIDIA launches Isaac SDK with simulation capabilities. Academic work shows transfer successful on object sorting tasks, with training times measured in minutes of simulation. CoRL 2018 introduces SPOTA algorithm for robust policy optimization via domain randomization.

• 2019: Field advances to routine transfer across manipulation (100% success on block-stacking/row-making), locomotion (biped and quadruped), and soft robotics. Domain randomization techniques mature: active search methods (CoRL 2019) and gradient-based parameter learning replace uniform sampling. Critical findings emerge: simulators show significant accuracy gaps (ICRA 2019), and embodied navigation agents exploit imperfections rather than learning genuine transfer (Georgia Tech/Facebook AI Research). Consensus: transfer is possible but reality gap remains; most deployments limited to controlled lab tasks.

• 2020: Consolidation to engineering practice with expanded platform availability. NVIDIA Isaac Sim 2020.1 establishes as industry standard. Real-world deployments expand: directional semantic grasping validated by NVIDIA research; quadrupedal locomotion succeeds on uneven terrain with dynamics randomization; task-oriented exploration outperforms passive methods in pouring and manipulation. Methodological advances: Bayesian Domain Randomization reduces prior knowledge requirements; active parameter search methods mature. Critical limitations documented: precision agriculture and other high-fidelity domains remain out of reach for current simulation techniques. Field consensus: sim-to-real is now proven for scoped manipulation and locomotion, but practitioners must carefully select applications within established capability boundaries.

• 2021: Ecosystem maturation with major vendor investments and methodological deepening. Meta AI launches Habitat 2.0 with 1,200 SPS performance (850× faster than existing platforms), expanding simulation-based training to home assistant robotics. MIT introduces PlasticineLab for deformable object manipulation, extending sim-to-real beyond rigid bodies. Domain randomization consolidates as standard practice; comprehensive review papers (Salvato survey, Robot Learning from Randomized Simulations) synthesize techniques and identify parameter selection as critical bottleneck. Practitioner adoption accelerates: Google researcher confirms extensive internal deployment across locomotion, navigation, and manipulation while emphasizing simulation necessity due to hardware wear constraints. Applied research shows mixed results: soft continuum arm visual servoing achieves 99.8% sim success but only 67% real-world zero-shot transfer, illustrating persistent reality gap in novel morphologies. Academic consensus: sim-to-real is a mature engineering discipline for scoped tasks, but generalization and high-precision domains remain open challenges.

• 2022-H1: Field consolidation with focus on specialized domain applications and precision manufacturing. Comprehensive review synthesizing domain randomization methodology published across leading institutions (Frontiers in Robotics and AI). Precision manufacturing validates sim-to-real for high-tolerance tasks: AIST research achieves 86% success on real-world insertion with ±0.01mm accuracy using curriculum learning plus domain randomization. Medical robotics demonstrates transfer effectiveness in constrained high-fidelity imaging: University of Toronto achieves 100% sim-to-real success on da Vinci surgical endoscope tasks. Locomotion advances continue: ROBOTIS-OP3 humanoid successfully transfers to uneven terrain and disturbances without force/torque sensing. Ecosystem maturity confirmed through independent comparative analysis of leading platforms (Gazebo vs NVIDIA Isaac Sim). Data augmentation techniques accelerate learning: University of Michigan shows 40% sim improvement and near-doubling of real physical success on deformable object tasks. By mid-2022, sim-to-real is established as an engineering practice with demonstrated capability across manipulation, locomotion, surgical robotics, and deformable objects—though adoption remains concentrated in research labs and early-stage commercial deployments rather than broad industry rollout.

• 2022-H2: Real-world deployment expansion and critical limitations surface simultaneously. Google demonstrates iterative sim-to-real transfer in human-robot table tennis with 150-hit rallies; dexterous manipulation competition (Real Robot Challenge 2022) highlights bridging RL and robotics via sim-to-real; object detection achieves 97.38% mAP using purely synthetic domain-randomized training. Commercial ecosystem matures: NVIDIA Isaac Sim and Omniverse see broad adoption (Amazon, PepsiCo scaling warehouse training). However, alternative approaches emerge: UC Berkeley's DayDreamer world-model trains robots directly on real hardware without simulators, achieving competitive performance and challenging simulation necessity assumptions. Research documents persistent reality gaps in swarm robotics and high-precision domains. By year-end 2022, consensus crystallizes: sim-to-real is a proven but specialized practice, effective within carefully defined capability boundaries but not universal; methodological focus shifts to automating domain randomization parameter selection and understanding when real-world learning outperforms simulation.

• 2023-H1: Industrial deployment expansion and methodological divergence. Trimble's Spot deployment for door detection exemplifies commercial adoption with Isaac Sim domain randomization (5% → 87% AP improvement). Visual manipulation policies mature with reduced real-world data requirements; photorealistic digital twins (3D Gaussian Splatting) enable zero-shot navigation transfer. Research challenges core assumptions: empirical studies show lower-fidelity simulators sometimes outperform high-fidelity variants by reducing overfitting and enabling faster training—questioning the universality of existing methodologies. Domain randomization consolidates as standard practice for scoped applications while simulation necessity itself becomes an open research question.

• 2023-H2: Mature deployment consolidation with rising alternative frameworks. NVIDIA DeXtreme achieves 42 years equivalent of real-world robot hand training in 32 hours of simulation (Isaac Gym + domain randomization); real-world transfer succeeds even with hardware malfunctions. Precision contact-rich manipulation (assembly, pivoting, screwing) demonstrates hybrid offline-online transfer; visual perception training via synthetic data reaches 93% success on real robots. However, critical challenge emerges: online DRL training with safety supervisors on physical vehicles demonstrates comparable performance to pure simulation approaches while bypassing the sim-to-real gap entirely (TUM/ICAR 2023). By year-end, field consensus shifts: sim-to-real is proven and reliable for scoped tasks but not universal; simulation necessity itself becomes contested. Practitioners view simulation as one tool among several, effective within defined boundaries but no longer assumed as default.

• 2024-Q1: Early industrial adoption expands with methodological advances in transfer automation. Toyota's production deployment with READY Robotics demonstrates sim-to-real integration in aluminum hot forging manufacturing via NVIDIA Isaac Sim and Omniverse integration. Research focus shifts toward reducing manual domain randomization design: DrEureka (RSS 2024) automates reward and randomization parameters via LLMs; TRANSIC (CoRL 2024) introduces human-in-the-loop framework achieving superior performance on contact-rich manipulation compared to baseline domain randomization. Real-to-sim-to-real approaches mature: RialTo pipeline robustifies imitation learning via digital twin refinement. Community infrastructure grows: SimplerEnv toolkit (CoRL 2024) provides standardized sim-based evaluation framework. Independent platform evaluations (NVIDIA Isaac systems) document continued refinement of simulation fidelity, though platform gaps remain. By Q1 2024, landscape reflects maturing industrial adoption combined with methodological focus on automating parameter tuning and expanding contact-rich task capability.

• 2024-Q2: Ecosystem consolidation with expanded real-world validation across domains. NVIDIA Isaac Lab enables training Boston Dynamics Spot locomotion at 85,000-95,000 FPS with zero-shot transfer to hardware; NVIDIA Isaac Perceptor and Manipulator workflows achieve GA with named adoption from ArcBest, BYD Electronics, KION, and Teradyne Robotics. Academic validation expands: Nature publication from NCSU/University of Michigan demonstrates physics-informed sim-to-real exoskeleton training producing 24.3% metabolic savings in human subjects; University of Alberta/Tokyo/NVIDIA publish industrial benchmark across eight manipulation tasks; IEEE Access papers validate transfer for autonomous robot path planning. Specialized domains advance: soft robot rehabilitation demonstrates data-imbalanced sim-to-real with 41-56% error reduction. By mid-2024, sim-to-real is established as mature engineering practice with expanding real-world deployments across locomotion, manipulation, and humanoid domains; ecosystem maturity confirmed through production deployments and broadening industrial adoption beyond research labs.

• 2024-Q3: Maturation and broadening of simulation ecosystem with strong vendor investment and specialized domain advances. NVIDIA announces expanded humanoid robotics tools at SIGGRAPH (NIM microservices, OSMO orchestration service) with developer program participation from 1X, Boston Dynamics, Figure, Fourier, and Skild AI, signaling major ecosystem growth. AutoMate framework from USC/NVIDIA achieves 84.5% mean success rate on real-world zero-shot assembly transfer across 20 different assemblies—demonstrating sim-to-real viability for contact-rich high-precision tasks. Robotic Sim2Real Competition (ICRA 2024) competition demonstrates consistent performance across simulation and real-world navigation/grasping/stacking tasks, showing technological readiness at engineering scale. Research expands beyond rigid-body control: ETH Zurich develops learned residual physics framework improving soft robot simulation accuracy by 60%, reducing reality gap for deformable systems; NeurIPS 2024 presents polynomial sample complexity bounds on simulation-guided exploration for RL domains where direct sim2real transfer fails. Industry adoption signals strong growth: robotic simulator market forecasted USD 1.89B expansion at 23.3% CAGR through 2028, with companies like NEURA Robotics joining developer programs targeting production scale (5 million robots by 2030). By quarter-end, sim-to-real landscape reflects consolidated industrial adoption with expanded capability boundaries beyond rigid manipulation, specialization in humanoid systems, and rising market investment in simulation infrastructure.

• 2024-Q4: Industrial production deployment expansion with continued research into methodology and limitations. Techman Robot (Quanta subsidiary) deploys NVIDIA Isaac Sim for electronics quality inspection, achieving 20% cycle time reduction and 70% programming time savings—exemplifying production-scale benefits. MIT CSAIL introduces LucidSim, a generative AI-based simulator that trains robot policies with 88% success rate on complex parkour tasks vs 15% for expert-only baselines, claiming superiority over domain randomization. Real-to-sim-to-real methods mature: MIT's RialTo pipeline uses 3D phone scans to build digital twins for RL, achieving 67% robustness improvement though exposing simulation gaming as an emergent challenge. Simultaneously, critical limitations emerge: meta-analysis of generative robotic simulators documents widespread low generalization capability, and NeurIPS 2024 research proposes indirect transfer (learning exploration policies in simulation when direct transfer fails), signaling fundamental unresolved gaps. NVIDIA's ROSCon announcements reinforce sim-first development paradigm with ecosystem partnerships (Universal Robots, Miso Robotics, Wheel.me). By quarter-end, sim-to-real is consolidated as a core industrial practice for factory automation and hardware programming acceleration, with expanding ecosystem and concrete production deployments. However, the field increasingly acknowledges that despite advances in methodology and hardened toolchains, the core reality gap persists—methodological diversity (domain randomization, generative approaches, real-to-sim-to-real) signals neither universal solution nor closure of fundamental challenges.

• 2025-Q1: Methodological consolidation with empirical validation of hybrid and advanced techniques. Comprehensive survey published on sim-to-real methods integrating foundation models, synthesizing formal taxonomies and maintaining literature repositories—confirming field maturity. Novel approach emerges: mixed sim-and-real co-training demonstrates 38% average real-world performance improvement across robot arm and humanoid tasks, directly addressing reality gap via hybrid training. Practical deployments continue: zero-shot transfer from NVIDIA Isaac Sim to real mobile robots achieves comparable performance to industry-standard ROS Nav2 navigation stack; ANYbotics integrates Surrealist simulation-based test generation framework into ANYmal quadruped workflow, successfully benchmarking algorithms though revealing significant weaknesses (40.3%-71.2% success rates). Critical research simultaneously quantifies reality gap magnitude: UAV controller transfer shows 100% more deviation in real flights vs simulation (though within 2m safe limits); ICLR 2025 workshops document that dynamic whole-body control behaviors (athletic loco-manipulation) remain challenging, requiring complex actuator modeling and pre-training strategies for transfer. By quarter-end, the landscape reflects consolidated engineering practice with expanding methodological toolkit (domain randomization, co-training, test generation) but continued acknowledgment that dynamic behaviors, contact-rich tasks, and high-precision domains still demand careful domain-specific adaptation—simulation necessity remains contested for novel morphologies.

• 2025-Q2: Advanced methodologies and real-world humanoid deployment accelerate. Novel real-to-sim-to-real approaches emerge: X-Sim framework learns manipulation policies from human videos, achieving 30% task progress improvement across embodiments with 10x data efficiency gains; Real-is-Sim uses dynamic digital twin synchronized at 60Hz to shift reality gap responsibility from policy to synchronization mechanisms. Humanoid robotics demonstrates practical impact: Robot Trains Robot framework enables ToddlerBot to double zero-shot walking speed within 20 minutes of real-world training using safe robotic guidance, achieving swing-up learning from scratch. Contact-rich manufacturing advances: Autodesk's dynamic compliance tuning framework enables zero-shot transfer on precision insertion tasks with sub-0.1mm clearances using RL-trained force planning. Academic research documents persistent challenges: comprehensive survey synthesizes gap causes and solutions while acknowledging closing the reality gap remains "one of the most pressing challenges"; latent-space analysis reveals direct policy transfer failures correlated with dynamics gap magnitude, demonstrating that simulation-trained approaches require careful parameter tuning and often underperform on novel domains. By mid-year, field consolidates around three dominant paradigms—pure simulation training with domain randomization, real-world adaptation with safety supervision, and hybrid co-training—with practitioners selecting based on task constraints and hardware availability; simulation necessity remains contested even as methodological diversity demonstrates field maturity.

• 2025-Q3: Methodological advances in perception and learning acceleration with mounting critical assessment. Camera Depth Models (CDMs) achieve 73%+ zero-shot sim-to-real success on depth-only manipulation without real-world fine-tuning, with open-source implementations for multiple camera systems. MIT's PhysicsGen multiplies VR demonstrations into thousands of simulations, yielding 60% improvement on dexterous hand tasks and 30% on multi-arm collaboration. SimLauncher hybrid framework achieves near-perfect success across contact-rich tasks by combining simulation pre-training with real-world RL. NVIDIA announcements include NeRD learned dynamics model with <0.1% accumulated reward error and zero-shot Franka transfer, advancing perception accuracy. Simultaneously, critical assessments surface: academic surveys document fundamental sim-to-real gap sources (dynamics, contact, state estimation) with transfer reliability remaining "limited despite mitigation strategies"; practitioner analyses identify persistent barriers (data scarcity, contact fidelity, noise replication) and catastrophic failure risk from domain shift. By quarter-end, the landscape reflects proven methodological toolkit (domain randomization, hybrid co-training, real-to-sim-to-real) with expanding capability demonstrations. However, field consensus crystallizes around core limitation: sim-to-real remains a valuable but specialized engineering practice, most reliable for perception and navigation, increasingly capable in scoped manipulation, but fundamentally constrained by persistent reality gap for dynamic and contact-rich behaviors—no universal solution has emerged despite six years of sustained research investment.

• 2025-Q4: Ecosystem consolidation and academic codification of mature practice with persistent reality gap acknowledgment. Academic consensus crystallizes: comprehensive survey accepted for Annual Review of Control, Robotics, and Autonomous Systems 2026 from NVIDIA, University of Washington, and ETH Zurich systematically catalogs sim-to-real gap causes, solutions, and evaluation metrics while affirming that "closing this gap remains one of the most pressing challenges in robotics" despite recent advances. Novel architectural solutions emerge: "real-is-sim" framework using dynamic digital twins (Embodied Gaussians) enables continuous simulator-real world alignment by decoupling policy execution from hardware. Ecosystem acceleration continues: generative world models (Marble) integrated with Isaac Sim reduce environment setup time from weeks to hours, signaling infrastructure maturation. Production deployment patterns solidify: healthcare robotics tutorial demonstrates 93% synthetic data usage with real teleoperation for remaining 7%, deployed on physical SO-ARM101 hardware—showing hybrid sim-real training as standard practice. Critical assessment grounds deployment expectations: practitioner analysis quantifies 24-30% real-world performance degradation from high-fidelity simulators and 30-50% success loss from modest scene variation, documenting persistent reality gap magnitude. Practitioner deployment guides emphasize production pitfalls: common failure modes include overfitting to narrow simulators (overconstrained policies) and excessive randomization (conservative, slow policies). By quarter-end 2025, the field consensus consolidates: sim-to-real is a proven, increasingly accessible engineering discipline with expanding methodological toolkit (domain randomization, co-training, real-to-sim-to-real, generative environment creation) and accelerating ecosystem support. However, the core tension remains unresolved—the reality gap persists as a fundamental challenge, making simulation-based training one effective tool among several rather than a universal solution, with practitioners selecting based on task constraints, hardware availability, and acceptance of deployment uncertainties.

• 2026-Jan: SimHum framework advances sim-and-human co-training, achieving 40% improvement in data efficiency and 62.5% OOD success with minimal real-world data. Critical assessment surfaces: humanoid robotics research identifies sim-to-real failures as fundamental bottleneck, noting simulations cannot yet reproduce contact, balance, and timing. Domain randomization techniques validated on real industrial robots (Franka, UR5, Baxter) with 84% transfer success; applied to free-floating object pre-grasp policies. Wheeled Lab platform demonstrates robust sim-to-real on wheeled systems with up to 100% real-world lap completeness and 90% elevation transfer success. Vendor ecosystem continues: NVIDIA positions sim-first development as industry standard for virtual commissioning and faster iteration. Landscape consensus solidifies: simulation-based training is proven for perception, navigation, and scoped manipulation, with hybrid co-training and co-supervision emerging as complementary paradigms; reality gap persists as core challenge for dynamic and contact-rich behaviors across morphologies.

• 2026-Feb: MOSAIC framework advances humanoid motion tracking via rapid residual adaptation bridging sim-to-real gap; validated on real hardware with robust teleoperation under realistic latency and noise. Empirical study (100 real-world runs across three platforms) identifies robust design choices for online RL on physical robots, documenting that default algorithmic choices can harm transfer. World-Gymnast demonstrates RL finetuning in action-conditioned video world models, achieving 18x outperformance over supervised learning on Bridge robot. Voxos industry analysis documents core adoption barrier: sim-to-real gap causes severe reliability drops (95% lab to 60% deployment), limiting factory automation despite $4.44B market and 39% growth. Isaac ecosystem advances: SimReady physics-validated assets and Marble generative world models reduce setup from weeks to hours; Sim2Val framework combines real/simulated validation. Critical assessment surfaces: synthetic data alone insufficient (61% accuracy); requires domain randomization (91% accuracy) plus human-in-the-loop for production. Landscape reflects three consolidated paradigms—pure simulation training, real-world adaptation, hybrid co-training—with practitioners selecting based on task constraints and acceptance of persistent reality gap for dynamic behaviors.

• 2026-Mar/Apr: Industrial deployment maturation with methodological diversification. Agility Robotics documented six-month physics investigation on Digit humanoid (28-DOF), identifying core sim-to-real gaps (contact dynamics, collision geometry, actuator energy propagation) and resolving them through hardware-specific physics understanding rather than reward engineering. Toyota's humanoid deployment case study demonstrates practical sim-to-real challenges on walking and dribbling tasks, with domain randomization and Real2Sim calibration enabling deployment. Menlo Research deployed novel Processor-in-the-Loop approach (actual firmware running in simulation with realistic timing constraints), achieving zero-shot locomotion transfer on Asimov Legs platform. ABB-NVIDIA partnership documented Foxconn pilot using simulation with synthetic data for assembly robot training before production deployment. NVIDIA released Isaac Lab as a unified open-source framework (April 17) consolidating GPU-accelerated RL, imitation learning, and motion planning, while a production deployment at Siemens using Isaac Sim-trained HMND 01 humanoids achieved 90%+ pick success with a 7-month prototype development cycle. VLA model benchmarks (CALVIN, SimplerEnv, LIBERO, DROID) showed Physical Intelligence pi0/pi0.5 leading real-robot transfer. UC San Diego demonstrated a three-phase sim-to-real framework for dexterous manipulation using disturbance injection and tactile force adaptation; ComSim introduced compositional simulation for scalable real-world data generation. Practitioner assessment continued to identify production constraints (repeatability, endurance, safety) as the point where simulation-trained robots most commonly fail. By late April 2026, the consolidated landscape confirms sim-to-real as core industrial practice for select domains with expanding ecosystem and methodology toolkit, though the core reality gap persists as a fundamental barrier to universal adoption.

• 2026-May: Production deployment evidence strengthened on two fronts: the Siemens-NVIDIA-Humanoid HMND 01 achieved 90%+ task success at 60 totes/hour with a 7-month development cycle (vs. 18-24 months traditional), and NVIDIA + Universal Robots validated contact-rich UR10e gear insertion with multi-configuration real-world transfer, while ABB launched RobotStudio HyperReality with Omniverse (99% sim-accuracy) and AWS demonstrated 4,096-parallel RL training on a single GPU with 91.5% lift success. The persistent reality gap was simultaneously restated with authority: former NASA robotics chief quantified 90% simulation success reducing to 12% real-world success on uncontrolled tasks, while practitioner assessment confirmed the contact-rich manipulation gap has not narrowed in 5 years — leaving simulation-based training a high-value but carefully scoped tool rather than a universal default.

• 2026-Jun: ICRA 2026 consolidated sim-to-real as a production engineering decision: NVIDIA's COMPASS framework achieved 80% real-world navigation success from simulation-only training across 28 ICRA papers; Grasp-MPC reached 75% success on real robots trained on 2 million synthetic trajectories; SPARR assembly workflows showed 38% success-rate improvement; HyperSim demonstrated 80-95% success with 35% robustness improvement across 400 real-world executions. Critical assessment simultaneously quantified the remaining gap: 80% success is insufficient for safety-critical deployment where a 20% error rate is catastrophic, and contact dynamics, friction, and sensor noise remain notoriously difficult to simulate faithfully — leaving sim-to-real as a high-value but task-scoped tool requiring careful selection and typically hybrid real-world fine-tuning rather than zero-shot transfer.