Simulation-based robot training
218 evidence items
Training robots in simulated environments (sim-to-real transfer) before deploying learned behaviours in the physical world. Includes domain randomisation and physics simulation; distinct from digital twins which model existing processes rather than training new behaviours.
Overview
Simulation-based robot training teaches robots new behaviours in physics simulators, using domain randomisation to carry them into the physical world, rather than modelling processes that already exist. It matters because it turns scarce, slow real-world data collection into cheap, parallel iteration. Mature open tooling and measurable production deployments now make it viable for navigation, perception and gross manipulation. It is a leading-edge practice and steady because the path is clear only in those scoped domains. For contact-rich, dynamic and safety-critical tasks the reality gap persists, and simulation moves cost into building and validating assets rather than removing it. Independent analyst recognition has also never appeared. Until transfer works reliably beyond the tasks early movers choose, it rewards specialist teams rather than any competent one.
Current Landscape
GPU-accelerated simulation is now shared infrastructure for physical AI. NVIDIA Isaac Lab 3.0.0-EA runs PhysX and Newton physics backends in one open-source robot-learning framework. Isaac Sim 6.2.0 ships policy examples for both engines, but NVIDIA flags Newton support as experimental and warns that success on one backend does not guarantee identical behaviour on another. Hugging Face's LeRobot v0.6.0 added world-model policies, unified benchmarks and DAgger-loop corrections. LeRobot also released an Apache-2.0 SO-101 tabletop environment in MuJoCo, with an Isaac Sim twin that shares its cameras, observations and actions. Open recipes now reach hobby hardware: Hugging Face's $399 Microduck biped ships with policies trained by reinforcement learning.
Evaluation practice is tightening as transfer claims multiply. NVIDIA RoboLab offers robot-agnostic policy diagnostics with Clopper-Pearson confidence intervals. Voxel51 notes that nine successes in ten rollouts carries an exact 95% interval of about 56% to 99.7%. A VLABench audit found simulated results highly sensitive to configuration. SmolVLA and pi0 completed 0/210 and 1/210 tasks at a 6-second cutoff, rising to 73/210 and 101/210 at 20 seconds. The authors say their results do not show sim-to-real transfer.
Simulated benchmarks can rank policies correctly while still flattering them. The SO-101 Bench authors built an Isaac Lab digital twin of a low-cost arm. Five policies trained without real data kept their exact ranking when run on the physical arm in an unseen environment (Spearman rho=1.00). The same authors report that simulation overestimated success by 9 to 31 percentage points.
Research is pushing parallel training to a new scale. The Success Guided Sampling method concentrates training on tasks at the edge of a policy's ability across over one million parallel simulated environments. It solves multi-terrain quadruped locomotion and contact-rich NIST assembly tasks that prior methods fail. Its policies, distilled to RGB input, transferred zero-shot to several real assembly tasks.
Simulation-only data is starting to beat real demonstrations. SimVLA was trained entirely on synthetic data from procedurally generated kitchens. Its authors report that it outperforms policies trained on 50 in-domain real-world demonstrations. The Zero-HLM authors report that GR00T N1.7 trained only in simulation reached 66.7% mean full-task success on a Unitree G1, a 20-percentage-point gain over real-data training. Apple researchers report that HALO, a humanoid that never saw a real home in training, achieved 91% success across 360 trials in six unmapped real-world environments.
Scene generation is being automated to cut setup cost. MIT reports AI agents building virtual playgrounds for robot training data. An agentic, simulation-ready 3D world engine targets the same cost. Zero-HLM rebuilds smartphone-captured rooms as digital twins. It expands 45 teleoperated simulation demonstrations into 4,800 physically validated ones.
Training recipes increasingly combine simulated reinforcement learning with real data. The RL-Co authors warm-start vision-language-action policies on mixed real and simulated demonstrations. They then fine-tune with reinforcement learning in simulation, while a loss on real-world data anchors the policy. They report +24% real-world success on OpenVLA and +20% on a second model, compared with SFT-based sim-real co-training on four tabletop tasks.
Real-to-sim reconstruction is becoming the standard way to ground simulators. Grounded-Dreamer reconstructs objects from a few real images and keeps only world-model trajectories that succeed in simulation. Robot policies trained on that data, with no human demonstrations, were competitive with demonstration-trained policies across seven real tasks. NC AI is wiring live factory data from Hyundai, LG and Coupang into a world model.
Transfer is spreading beyond arms and legged robots. An ASV system-identification pipeline needed only a CAD model and brief open-water trajectories to deliver zero-shot path following and station keeping on a BlueBoat. A Scientific Reports study flew one simulation-trained PPO policy on a Crazyflie 2.1 carrying a suspended payload, without tuning. It stayed stable under 1.0 m/s crosswind, though the authors report degraded performance at higher velocities from aerodynamic drag and actuator saturation.
Contact-rich transfer is advancing through system identification rather than richer sensing. A proprioception-only peg-insertion policy was trained in simulation on an arm model whose armature, damping and friction were identified from real measurements. It seated 173 of 237 real trials (73%) across eight pegs, three of them outside the training distribution. Its companion generator, mjfit, builds MuJoCo assemblies.
Learned navigation is closing on production stacks. On Idealworks iw.hub robots, digital-twin training with NVIDIA Cosmos sim-to-real augmentation raised real-world success from 65.3% to 91.7% across four scenarios, compared with simulation-only training. That brought it within 8.3 points of the production navigation stack, running fully onboard. The trials ran in a test facility rather than a customer rollout, and the robot developer likely ran the evaluation itself.
Commercial deployments report shorter development cycles. Noble Machines reports cutting humanoid development from 4 years and 50 engineers to 18 months and 15 engineers using Isaac Sim and Isaac Lab. Siemens reports that the HMND 01 humanoid reached 90% task success in a 7-month development cycle, against 18-24 months traditionally. Qingcang Robotics reports deploying a L'Oreal line in 30 days with 99% success. Roborax reports a 91% average transfer rate across 41 centres, with each result checked by paired real-world validation.
Large manufacturers are building out simulation and data capacity. LG and NVIDIA are racing to build 100,000 hours of robot training data in Seoul. Techman Robot has moved Real-Sim-Real workflows from virtual validation onto production lines. These combine 3D reconstruction, Omniverse simulation and AI spatial positioning. Boston Dynamics trains Atlas for factory work at Hyundai with teleoperation and reinforcement learning in simulation.
Production reliability is now the threshold that matters. NVIDIA's Seattle Robotics Lab and Foxconn deployed Isaac Lab reinforcement-learning policies for connector insertion on GB300 tester trays. They report high sim-to-real success on the large connector. Their busbar assembly solution reached success rates above 95%, short of the 99.5% Foxconn requested. Its 160-second cycle missed a 124-second target. The authors call research norms of 50-60% success over 10-20 trials inadequate for production. They found a classical perception-planning-control stack sufficient for busbar assembly.
Measured reality gaps stay large outside tightly scoped tasks. Fortune 500 retailers documented 92%+ simulation success falling to 67-74% on live warehouse floors within two weeks. In a Shanghai Jiao Tong paper-cup pouring test, the ManipTrans baseline fell from 71.7% success in simulation to 0% on hardware. The authors attribute the drop to cup deformation and heavier payloads. A former NASA robotics chief warns that policies tuned to narrow simulation distributions fail on uncontrolled real-world tasks.
Domain randomisation itself is drawing criticism. A separate critical assessment argues that it hits fundamental limits in industrial contact-rich deployment. Voxel51 argues that blanket domain randomisation can produce a worse, overly conservative policy. It adds that tabletop manipulation remains hard where quadruped locomotion is largely considered solved.
Constraining policy structure is another way to narrow the gap. A Science Robotics study ran 250 trials on physical e-puck robots. A continuous-action baseline that earned among the highest simulation rewards suffered the largest performance drops on hardware. Policies limited to selecting predefined behaviour modules transferred more reliably. A review of 17 swarm-robotics studies finds that learned controllers transfer to hardware in some studies but show a large simulation-to-reality gap in others.
Hardware changes are being used to absorb the remaining gap. EPFL researchers trained a policy in rigid-body simulation and added mechanical compliance only at deployment. Under a +6 mm table-height perturbation, this lifted pick-and-lift success from 36.7% to 63.3%. Safety stops fell from 53.3% to 16.7%. The opposite trade-off draws criticism. Critics say 'Design for Simulation' strips capability from hardware such as the Unitree G1's joints to keep simulation tractable.
Practitioners argue that simulation moves cost rather than removing it. TaskUs' Rick Balzano proposes a '90–10 rule': whether finishing the last 10 percent after simulation takes 90 percent of the budget. He adds that advertised free-roaming factory robots worked in a 40-by-40 area with teleoperators on hand. Palatial makes the same argument about robot data cost. A CoreWeave tutorial shows sim-trained locomotion policies gaming simplified physics by vibrating their motors. It reports that domain randomisation makes training 2-3x longer.
Claims for foundation models trained on simulation stacks remain unverified. Skild AI reports 66% average per-step cumulative success for S1 on unseen tasks, against 9% for a VLA baseline. S1 was trained partly in Isaac Sim and Isaac Lab. Aivancity notes that the metric relied on human resets and that no independent replication existed. Until transfer holds on contact-rich, long-tail production tasks, deployment stays concentrated in navigation, perception and gross manipulation.
Tier History
Evidence (218)
— Apple researchers report a humanoid trained entirely in simulated procedural homes reaching 91% zero-shot success over 360 real trials in six unmapped environments; self-reported, no baselines shown.
— Negative signal: the ManipTrans baseline falls from 71.7% success in simulation to 0% on hardware in paper-cup pouring, which the authors attribute to cup deformation and heavier payloads.
— Isaac Lab twin of a low-cost arm: five policies trained without real data keep their exact ranking on hardware (Spearman rho=1.00), but simulation overestimates success by 9 to 31 points.
— Simulation-only data beats real teleoperation data on a Unitree G1: GR00T N1.7 reaches 66.7% full-task success, 20 points above real-data training, with pooled generalisation rising from 23% to 69%.
— One simulation-trained PPO policy flies a Crazyflie 2.1 with suspended payload without tuning, stable under 1.0 m/s crosswind, but degrades at higher speed from drag and actuator saturation.
213 more · latest 2026-10-23 →
— Negative signal: an independent review of 17 swarm studies finds learned controllers transfer to hardware in some studies but show a large simulation-to-reality gap in others.
— Negative signal: argues that transfer failures come from dynamics, perception or task mismatch, that blanket domain randomisation can produce worse policies, and that small rollout counts carry wide confidence intervals.
— Peer-reviewed study across 250 physical e-puck runs: the policy with the highest simulation reward lost most on hardware, and module-constrained policies transferred more reliably.
— Digital-twin training with Cosmos augmentation raised real Idealworks AMR navigation success from 65.3% to 91.7%, within 8.3 points of the production stack. Test facility only, and likely self-run.
— Negative signal on simulation-based evaluation: SmolVLA and pi0 completion swings from 0/210 and 1/210 to 73/210 and 101/210 depending on the time cutoff, and sim-to-real transfer is explicitly not shown.
— Contact-rich evidence: a sim-trained, proprioception-only insertion policy built on a system-identified arm seated 173 of 237 real trials (73%) across eight pegs. Also releases the mjfit MuJoCo generator.
— LeRobot releases an Apache-2.0 SO-101 MuJoCo training environment with a matching Isaac Sim twin, so the same policies run against both simulators. Tooling only, with no results.
— Isaac Sim 6.2.0 ships Isaac Lab RL policy examples on PhysX and Newton. It flags Newton as experimental and warns that cross-engine (sim-to-sim) transfer does not guarantee identical behaviour.
— A VLA trained only on synthetic simulation data beats policies trained on 50 in-domain real demonstrations in real kitchens and homes. Self-reported preprint without exact rates.
— Real-to-sim-grounded world-model data trains real robot policies with no human demonstrations. They are competitive with demo-trained policies across seven tasks and better out of distribution.
— Success Guided Sampling scales sim RL to over one million parallel environments and solves contact-rich NIST assembly. Its distilled RGB policies transferred zero-shot to real hardware.
— Critical analysis: Skild AI's S1, trained partly on Isaac Sim/Lab, claims 66% versus 9% per-step success, but the metric needed human resets and no independent replication existed.
— Extends sim-to-real RL to surface vessels: system identification from a CAD model and brief field trajectories gave zero-shot path following and station keeping on a BlueBoat ASV.
— NVIDIA and Foxconn deployed Isaac Lab sim-to-real RL for GB300 connector insertion, with candid production gaps: success above 95% against a 99.5% target, and a 160 s cycle against a 124 s target.
— Alpha, single-maintainer platform that drives MuJoCo/mjlab and LeRobot through a 77-tool MCP server. It says openly that no policy trained with it has run on a real robot yet.
— Negative signal: TaskUs VP poses a '90–10 rule' on post-simulation cost and says demoed free-roaming robots ran in a 40-by-40 area with teleoperators, arguing simulation alone cannot ready robots.
— Peer-reviewed EPFL work: compliance added at deployment lifted sim-trained pick-and-lift success from 36.7% to 63.3% and cut safety stops from 53.3% to 16.7%, a fix for the contact-rich last-centimetre gap.
— Major framework release adds a standalone Newton physics backend alongside PhysX, with domain randomisation, tiled rendering and bundled robot assets under a BSD-3 licence.
— Sim-trained dexterous policies reach 89.3% zero-shot success over 300 real-world trials on 30 objects, with residual RL in simulation making retargeted human motion physically feasible.
— Multi-task training in simulation gives 56% zero-shot insertion success on unseen objects, against 7% for a model-free baseline. The figures are in simulation, not on hardware.
— Limitation signal: contact dynamics, sensor noise, materials and lighting rarely transfer cleanly, and the bottleneck is moving downstream to real-time inference (Intel's 100ms target for π0.5).
— GA vendor platform covering URDF/MJCF/USD import, PhysX or Newton physics, Replicator synthetic data, and software-in-the-loop validation of ROS 2 stacks before hardware.
— Critical vendor opinion: asset building, validation and compute stay costly, and the sim-to-real gap starts with wrongly modelled joints and physics. Cites the GR00T N1 figure of 750k trajectories in 11 hours.
— Negative signal from named practitioners: teaching one industrial skill through teleoperation, world models and simulation takes six months, against the days or weeks automotive plants need.
— Hardware co-designed for simulation: its cable-transmission model and identified actuation map let sim-trained policies run with no fine-tuning. The authors report no transfer metrics.
— Production-ready $399 open-source biped with all behaviors (walking, kicking, self-recovery) trained via PPO in MuJoCo; ships with seven policies and open-source sim-to-real recipe including actuator modeling and domain randomization.
— Industry roundup identifying Amazon Robotics integration of NVIDIA physical AI infrastructure (Jetson, Omniverse, Isaac) with AWS 2M GPU deployment plans, signaling enterprise-scale sim-to-real adoption in logistics.
— Noble Machines deployed NVIDIA Isaac Sim and Isaac Lab to compress humanoid development from 4 years/50 engineers to 18 months/15 engineers, demonstrating 3× acceleration via simulation-first methodology and quantified timeline reduction.
— Techman Robot deployed Real-Sim-Real workflow combining 3D reconstruction, NVIDIA Omniverse simulation, and AI spatial positioning for manufacturing assembly, establishing production-ready deployment pipeline.
— LG-NVIDIA MOU targets 100,000 hours combined real+synthetic training data by end-2026, surpassing Ant Group's 60k-hour corpus; DataFactory spans 10,000m² with four training environments, signaling major OEM infrastructure investment.
— Commercial sim-to-real data service with 10K+ daily simulations and 91% average transfer rate via paired real-world validation across 41 centers in 12 countries, demonstrating ecosystem maturity and production viability.
— Peer-reviewed zero-shot sim-to-real transfer on Unitree G1 humanoid achieving 73.3% success on unseen object-pick tasks, trained entirely in simulation using procedurally-generated scenes and heavy domain randomization.
— Dexterous manipulation research quantifies deployment boundary: WUJI-Scissors achieved 98.7% sim success but 0/10 real hardware trials due to non-backdrivable motors and contact-rich dynamics mismatch—reveals fundamental sim-to-real gap in contact manipulation.
— Niantic Spatial + Flexion zero-shot humanoid navigation deployment using photorealistic Gaussian splatting reconstruction; RGB-only policy transfer without on-site fine-tuning adaptation demonstrates deterministic deployment pathway.
— CVPR 2026 peer-reviewed study: NVIDIA/CMU/Berkeley humanoid robot sim-to-real framework achieves 91.5% real-world success (54/59 trials) on Unitree G1 via reference state initialization and visual randomization at 64-GPU scale.
— Standardized skill-transfer-simulate-deploy pipeline validated at scale across OEMs: Figure AI 1,250+ operating hours at BMW Spartanburg, Mercedes-Benz/Apptronius, Hyundai/Boston Dynamics Atlas, Amazon/Agility Digit—humanoid training methodology consolidation.
— Alibaba DAMO production-readiness framework: four-stage validation pipeline with physics fidelity metrics (PSNR 35dB+, 95%+ behavior consistency), closed-loop feedback from field trials; establishes production methodology for scaled humanoid deployment.
— NVIDIA Isaac for Healthcare Medical Physics Simulation GA with named enterprise adoption (J&J MedTech, CMR Surgical, Medtronic) enables 8,192-parallel environment RL training in <2 minutes vs 5+ hours—production deployment in surgical robotics.
— Practitioner documentation of production R2S2R framework operating since May 2023 across 60+ customer sites with 238+ machine types, cataloguing 18,000+ V1-V3 edge cases (real deployment scenarios) in open Smiling Buddha library for foundation model training.
— Named partnership implementing bidirectional development loop where deployed robots at Hyundai, LG, and Coupang feed live factory data into world model training; achieved 70-80% SOTA performance at 25% GPU cost with 10K hours synthetic video generated in 11 days.
— Production release of Isaac for Healthcare Medical Physics Simulation addressing healthcare robotics data scarcity via GPU-native modular framework supporting classical physics solvers and generative world models for synthetic RL policy training.
— Expert assessment identifying fundamental sim-to-real barriers: domain randomization cannot bridge unmodeled phenomena (micro-slip, thermal expansion, surface contamination); teams should budget 40-60% project timeline for sim-to-real debugging versus common 10-20% assumptions.
— Research benchmark demonstrating 0.92 Pearson correlation between simulation and real-world policy performance via 1,000+ physically-based rendering 3D assets, validating visual fidelity as predictor of sim-to-real transfer success.
— Niantic Spatial, Flexion, and NVIDIA demonstrated end-to-end real-to-sim-to-real pipeline: site RGB reconstruction → photorealistic Gaussian splatting rendering → Isaac Sim RL training → zero-shot humanoid navigation transfer to real offices in hours.
— Novel framework automating real-to-sim conversion via vision-language agents, reducing manual reconstruction bottleneck across rigid-object, deformable-object, and humanoid domains—enabling scalable digital twin generation for RL training.
— NVIDIA's production GA of Cosmos 3 Edge (4B-parameter foundation model) for on-device physics simulation and closed-loop synthetic data generation, enabling faster robot training iteration with reduced computational overhead.
— Novel sim-to-real method addressing partial observability by extracting unobservable dynamics backward from observed transitions, with validation on humanoid, quadruped, and manipulator platforms including Go2 real-robot deployment.
— Analysis of world-action models achieving 65-95% sim-to-real transfer; demonstrates shift from data-scaling to model-architecture problem; multiple named systems (Cosmos, Fysics, HyperSim) with transfer metrics.
— NVIDIA simulation benchmarking platform addressing reproducibility gaps: robot-agnostic task generation, diagnostic failure analytics, statistical rigor via Clopper-Pearson intervals; integrating into Isaac Lab August 2026.
— MIT SceneSmith system generated 1,300+ diverse simulation scenes with 6x more objects per scene than prior methods; agents validated realism >99% agreement with humans—automated environment creation for sim training.
— EmbodiedGen V2 generative 3D world engine for sim-to-real: online RL improved simulation task success 9.7%→79.8%, real-robot transfer 21.7%→75.0%; automated environment creation at scale.
— Empirical quantification of sim-to-real failure: simulation-trained agent degraded 1160% vs human-level performance after real-world deployment; documents fundamental transfer unreliability on uncontrolled tasks.
— Atlas humanoid trained on mocap-to-sim RL pipeline in hours vs prior weeks; real-world deployment at stadium-scale public venue demonstrating training acceleration on complex behaviors.
— 8B model trained on 400k simulated trajectories across 6k scenes; achieved 76.6% real-world success on unseen environments without real-world data collection—validated sim-only training at scale.
— Official NVIDIA Isaac Lab sim-to-real workflow: UR10e gear insertion policy trained in simulation with domain randomization, deployed on real hardware with three-stage consistency framework.
— Full production-line deployment: sim-to-real training accelerated L'Oreal cosmetics line from 3-6 months to 30 days with >99% stable success via VLA + digital twin validation.
— Critical assessment: Fortune 500 retailers (Amazon, Walmart, Target, Ocado) reported 92%+ simulation success degrading to 67-74% live deployment within 2 weeks; named failure breakdown (41% navigation, 33% grasping, 26% collision).
— Major open-source framework release: world-model policies, unified simulation benchmarks, reward models, and closed-loop DAgger-style correction loop enabling production-ready sim-based training.
— First peer-reviewed demonstration of zero-shot sim-to-real for multi-finger dexterous hands using tactile simulation and actuator dynamics modeling; dexterous manipulation breakthrough.
— JD Tech + Tsinghua/Peking/Beihang white paper detailing cloud-native simulation infrastructure for embodied intelligence training at scale, with phased roadmap addressing sim-to-real gap mitigation.
— Practitioner analysis documenting Boston Dynamics zero-shot sim-to-real transfer: Atlas generalizing from 23-32kg training to 45kg load, with KB Securities analyst claim of millions of training hours per day in simulation.
— Critical assessment from 40-year simulation veteran arguing Design-for-Simulation philosophy prioritizes software convenience over machine performance, with hardware capability examples (Unitree G1 joint redesign, tendon-hand simplification).
— Industry analysis of 2026 humanoid deployments showing simulation+teleoperation as dominant training paradigm; identifies sim-to-real gap as central constraint for production-grade reliability (99%+ required vs. 80-90% achieved).
— CVPR 2026 research on continual sim-to-real adaptation using geometry-aware mixture-of-experts, achieving 52% success-rate improvement on real manipulation tasks with 6x data efficiency.
— Deployment outcome verified: humanoid walking stable on real hardware within hours of deployment using 7-day Isaac Lab training, with explicit acknowledgment that locomotion is simplest task and contact dynamics remain unresolved.
— CBS 60 Minutes deployment documentation: Atlas at Hyundai using VR teleoperation with supervised learning and RL in simulation (4,000+ digital Atlas training), successful skill transfer to real hardware, zero-shot deployment.
— Peer-reviewed research achieving zero-shot deformable object transfer (80%+ success on plastic bags and silk grasping) without teleoperation, surpassing real-data baselines under visual distribution shifts.
— Novel SPAD methodology swapping physics simulation at deployment for learned real-world state predictor, narrowing sim-to-real gap on high-speed dynamic task with successful real-world table tennis transfer.
— Negative-leaning tutorial: sim-trained locomotion policies exploit simplified physics (100 Hz motor vibration), and domain randomisation costs 2-3x longer training while cutting real-world drop to ~20%.
— COMPASS framework achieved 80% real-world navigation success from simulation-only training; 28 ICRA papers validating methodology positions sim-to-real as engineering decision.
— Critical assessment: 80% success is insufficient for safety-critical deployment; long-tail failures and 20% error rate make current simulation-trained policies unsuitable for production without additional validation.
— Eight ICRA 2026 papers: COMPASS 80% navigation, Grasp-MPC 75% novel objects, SPARR 38% assembly improvement, demonstrating production-grade sim-to-real across navigation, grasping, assembly, manipulation.
— ICRA 2026 vendor results: GPU-accelerated sim-to-real with quantified metrics across grasping, navigation, assembly; demonstrates ecosystem maturity for production deployment.
— Peer-reviewed empirical study: 400 real-world executions across two manipulation models; 80-95% success rates, 35% robustness improvement under perturbation via synthetic data, adversarial generation, and co-training.
— Practitioner perspective from Formic Robotics CTO: sim-to-real gap is structural, cannot be eliminated, only made robust to; contact dynamics and sensor modeling remain persistent bottlenecks.
— Expert critical assessment quantifying persistent reality gap: 90% simulation success → 12% real-world success; argues current optimization for controlled demonstrations creates brittle systems.
— Siemens-NVIDIA-Humanoid HMND 01 production deployment with Isaac Sim and Isaac Lab; 7-month development cycle (vs. 18-24 months traditional); 90%+ task success, 60 totes/hour industrial metrics.
— NVIDIA + Universal Robots UR10e gear insertion (contact-rich manipulation); PPO with domain randomization; proven multi-configuration real-world validation exemplifying modern sim-to-real manufacturing workflow.
— AWS case study: 4,096 parallel simulated environments on single L4 GPU using PPO; suction-cup gripper task achieved 91.5% lift success and 77.9% placement accuracy; demonstrates modern large-scale sim-based training infrastructure.
— Tier-1 vendor product launch: RobotStudio HyperReality integrating Omniverse for synthetic data generation; 99% simulation accuracy; Foxconn and WORKR pilots indicate production-ready ecosystem.
— NVIDIA GR00T-Dreams synthetic motion pipeline yields ~40% improvement over real-world-only baselines; comprehensive ecosystem analysis of 2026 humanoid deployments validating simulation-training scaling.
— Critical assessment across 1,400+ Unitree G1 episodes documenting persistent gaps: 1.1x task success, 1.5x manipulation accuracy, 50x grasp adaptiveness—exposing fundamental simulation limitations.
— Balanced practitioner guide noting sim-to-real gap in contact-rich manipulation has not narrowed in 5 years; identifies real data as essential for all commercial deployments despite synthetic progress.
— IEEE Transactions on Field Robotics: ASV deployment with centimeter-level accuracy, honest identification of failure modes (insufficient actuation-model fidelity), and practical mitigation strategies.
— ETH Zurich's production-ready framework demonstrating sim-to-real transfer across multiple platforms (ANYmal, Unitree A1, Cassie) with domain randomization, proving practical deployment at scale.
— VLM-guided domain randomization with tactile-visual fusion achieves 78.2% real-world success on contact-rich manipulation, reducing sim-to-real gap to 8.3% through advanced methodology.
— Multi-vendor production deployments showing 99% sim-accuracy, 50% reduction in product cycles, 80% commissioning time reduction, and measured synthetic-data sufficiency for production-grade AI training.
— Simulation-first training compressed prototype development from 18-24 months to 7 months; deployed at Siemens achieved 90% pick-and-place success, 60 tote moves/hour in production.
— RSS 2026 peer-reviewed research demonstrating zero-shot sim-to-real transfer on underactuated humanoid ballbot via friction-aware RL framework with asymmetric actor-critic training.
— Benchmarking VLA models trained in simulation (CALVIN, SimplerEnv, LIBERO, DROID) showing Physical Intelligence pi0/pi0.5 leading real-robot transfer across manipulation tasks.
— Named deployment showing simulation-driven development: Isaac Sim reduced cycle time 7-24 months; live trial at Siemens achieved 90%+ task success on logistics tote handling.
— Official NVIDIA documentation positioning Isaac Lab as a mature, open-source framework for simulation-based robot learning, supporting RL, imitation learning, and motion planning with GPU acceleration.
— Real production deployment of HMND 01 humanoid at Siemens facility using Isaac Sim simulation-first training, achieving 90% pick success and 7-month prototype development.
— UC San Diego research demonstrating three-phase sim-to-real framework for dexterous manipulation: simulation with disturbance injection, policy distillation, and tactile force adaptation.
— Comprehensive 33-milestone deployment tracker showing Agility Digit as only humanoid achieving full commercial cycle with specific customer counts and operational metrics.
— Novel hybrid approach combining classical and neural simulation to generate training data with real-world consistency, enabling closed-loop sim-real data augmentation pipeline.
— Critical assessment of sim-to-real limitations: robots flawless in demonstrations fail under production constraints (repeatability, endurance, safety). Identifies core barrier to scaling.
— Real-world deployment tracking with specific companies and metrics showing outcomes of simulation-based training approaches—Agility Robotics with GXO (100,000+ totes), Figure AI with BMW pilot.
— Toyota detailed case study on humanoid robot deployment (walking, basketball dribbling) showing concrete sim-to-real challenges and mitigation strategies. Domain randomization and Real2Sim calibration enabled practical deployment after iterative refinement.
— Menlo Research deployed novel Processor-in-the-Loop approach on Asimov Legs: actual robot firmware runs in simulation with realistic timing/comms constraints. Key insight: sim-to-real failures stem from embedded system realities (CAN delays, thread misses, IMU drift), not physics gaps. Zero-shot sim-to-real locomotion achieved (forward, backward, lateral walk, push recovery).
— Novel simulation-based training framework (OmniReset) using programmatically diverse resets to enable zero-shot sim-to-real transfer for complex dexterous manipulation without curricula or demonstrations—exemplifies leading-edge practice advancement.
— Named company (Agility Robotics) deployed NVIDIA Isaac Lab for Digit humanoid (28-DOF, actively balancing). Systematic 6-month investigation identified specific sim-to-real gaps: toe-impact contact dynamics, collision geometry, actuator energy propagation, closed-chain kinematics. Rather than stacking rewards, they fixed underlying physics model gaps—production-level problem-solving on complex dynamic platform.
— Practitioner analysis identifying core failure modes of sim-to-real transfer in real deployments: fine motor control, adaptation to unexpected changes, and domain mismatch.
— Named pilot by Foxconn (world's largest electronics manufacturer) using simulation with synthetic data to train assembly robots before production deployment.
— ICLR 2026 peer-reviewed research establishing theoretical foundations for offline domain randomization (ODR), directly addressing how to leverage real-world data to improve simulator parameter distributions for better sim-to-real transfer.
— Large-sample empirical study of 100 real-world training runs across three robotic platforms identifying robust design choices for successful online RL on physical hardware; documents that default algorithmic choices can be harmful to sim-to-real transfer.
— Critical assessment from robotics annotation provider: synthetic-only approaches achieve 61% sim-to-real accuracy without domain randomization, rising to 91% with proper randomization but requiring human-in-the-loop review; documents necessity of combined pipelines for production reliability.
— Open-source humanoid motion tracking system using rapid residual adaptation to bridge sim-to-real gap; validated on real hardware with robust offline replay and online teleoperation under latency and noise.
— Industry analysis documents that sim-to-real gap causes severe reliability drops (95% lab success to 60% deployment); identifies core adoption barrier preventing wider factory automation adoption despite 39% market growth and $4.44B market size.
— Ecosystem advancement: SimReady physics-validated assets loadable directly into Isaac Sim/Lab; generative world models (Marble) reduce environment setup from weeks to hours; Sim2Val framework combines real/simulated tests to reduce physical mileage requirements.
— RL finetuning in action-conditioned video world models for vision-language-action policies; demonstrates 18x outperformance over supervised learning and 2x over software simulation on Bridge robot setup with cloud-based training capabilities.
— RL fine-tuning in simulation anchored by a real-data loss beats SFT sim-real co-training by +24% real-world success on OpenVLA and +20% on a second VLA across four tabletop tasks.
— SimHum framework combines synthetic simulation data with real-world human observations, achieving 40% improvement in data efficiency and 62.5% out-of-distribution success—advancing hybrid sim-and-real co-training for robotic manipulation.
— NVIDIA's production offering for robot learning in simulation, emphasizing sim-first approach enabling parallel training across robot embodiments (AMRs, arms, humanoids) with synthetic and real-world data integration.
— Critical assessment identifying sim-to-real failures as fundamental bottleneck for humanoid robotics; simulations cannot yet reproduce contact, balance, and timing dynamics needed for reliable movement, causing sharp performance drops outside controlled environments.
— Open-source platform integrating modular hardware, NVIDIA Isaac Lab high-fidelity simulation, and ROS for reproducible sim-to-real research. Demonstrates robust real-world transfer: 100% lap completeness on drift tasks, 90% elevation traversal success, 60% vision navigation success.
— Study on 7,000+ grasp trajectories validated across real robots (Baxter, Franka, UR5) shows domain randomization-based metrics strongly correlate with real-world transfer success; prioritizing candidates by DR metrics achieves 84% transfer success on Franka.
— Research on domain randomization for pre-grasp policies of free-floating objects shows policies trained in simulation with DR demonstrate consistent real-world behavior; integration with tactile sensors improves success rates on 6-DOF industrial robot.
— Practitioner-focused deployment guide to dynamics randomization for production automation. Outlines practical recipe: real robot measurements → parameter randomization → curriculum learning → reality checks → staged rollout. Documents common failure modes (overfitting to narrow simulator vs policy conservatism from wide randomization). Emphasizes implementation pitfalls for U.S. automation context.
— NVIDIA technical tutorial demonstrating integration of generative world models (Marble) with Isaac Sim to create photorealistic simulation-ready 3D environments from text prompts. Generative AI reduces environment setup time from weeks to hours. Shows ongoing ecosystem development accelerating simulation creation for robotics workflows.
— Research paper presenting 'real-is-sim' framework using dynamic digital twin (Embodied Gaussians) for continuous simulator-real world alignment. Policies run within simulator while real robot tracks simulated joints, decoupling policy execution from hardware. Validation on PushT task demonstrates strong correlation between simulated and real-world success rates. Novel architectural approach to reality gap mitigation.
— Critical practitioner analysis quantifying real-world performance degradation: 24-30% drop when policies transfer directly from high-fidelity simulators. Documents that insufficient scene variation causes 30-50% success reductions under modest perturbations. Identifies limitations in current benchmarks emphasizing scalability over fidelity. Provides grounded assessment of sim-to-real gap magnitude.
— Technical tutorial for surgical assistant robots using NVIDIA Isaac for Healthcare v0.4 with mixed sim-real training. Demonstrates 93% of policy training data generated synthetically in simulation with remainder from real teleoperation. Deployed on physical SO-ARM101 hardware. Shows production-scale mixed simulation approach in healthcare robotics.
— Comprehensive survey accepted for Annual Review of Control, Robotics, and Autonomous Systems 2026 by researchers from NVIDIA, University of Washington, and ETH Zurich. Systematically analyzes sim-to-real gap causes, solutions (domain randomization, real-to-sim transfer, co-training), and evaluation metrics. Notes that closing the gap remains a pressing challenge despite advances. Provides critical academic assessment of field maturity.
— NVIDIA Q3 2025 announcement of NeRD learned dynamics model achieving less than 0.1% accumulated reward error and zero-shot sim-to-real transfer on Franka manipulation, plus Dexplore and VT-Refine advances.
— Practitioner critique documenting persistent sim2real barriers: limited data scale, fidelity-contact-noise challenges, and catastrophic failure risk from domain shift—assessing why simulation-trained-only approaches remain unreliable.
— Academic survey analyzing sources of sim-to-real gap in bipedal locomotion (dynamics, contact modeling, state estimation), documenting that transfer reliability remains fundamentally limited despite existing mitigation strategies.
— Independent research introducing Camera Depth Models for zero-shot sim-to-real transfer, achieving 73%+ success on depth-based manipulation without real-world fine-tuning, with open-source toolkit for 5 camera types.
— MIT CSAIL PhysicsGen pipeline multiplies VR demonstrations into thousands of simulations, achieving 60% accuracy improvement on robotic hand manipulation and 30% improvement on virtual arm collaboration tasks.
— Hybrid framework combining simulation pre-training with real-world RL, achieving near-perfect success rates across contact-rich and dexterous hand tasks while improving sample efficiency over pure real-world learning.
— Research quantifying sim-to-real transfer failures in model-based RL: direct policy transfer shows performance degradation correlated with dynamics gap, demonstrating persistent challenges despite methodological advances.
— Comprehensive survey analyzing sim-to-real gap causes (dynamics, perception, actuation) and solutions (domain randomization, real-to-sim transfer, co-training), synthesizing empirical findings and identifying remaining open problems.
— Real-to-sim-to-real framework learning policies from human videos achieving 30% task progress improvement over baselines across 5 manipulation tasks with 10x data efficiency gains.
— Contact-rich industrial insertion framework using RL-trained force planner with dynamic compliance gains, enabling zero-shot real-world transfer on high-precision tasks with sub-0.1mm clearances.
— Real-world humanoid deployment using robot-teacher framework achieving 2x zero-shot walking speed improvement and 15-minute swing-up learning on ToddlerBot, demonstrating effective real-world policy adaptation.
— Dynamic digital twin framework synchronizing with real robot at 60Hz, shifting sim-to-real gap responsibility to digital twin synchronization and demonstrating consistent virtual-real results on long-horizon manipulation.
— Research from NVIDIA and academic institutions demonstrating mixed simulation-real data training improves real-world manipulation performance by 38% on average across robot arm and humanoid domains.
— Industrial adoption of Surrealist simulation-based test generation framework by ANYbotics for ANYmal quadruped, revealing significant algorithmic weaknesses (40.3% success) alongside superior alternatives (71.2%) in pilot evaluation.
— Peer-reviewed survey providing formal taxonomy of sim-to-real techniques and evaluating integration of foundation models, synthesizing consolidated methodological understanding while acknowledging persistent sim-to-real gap challenges.
— Research quantifying sim-to-real gap in UAV controller transfer: 100% more deviation in real flights vs simulation, though within 2m safe limits, documenting concrete magnitude of reality gap in aerial control.
— Research demonstrating zero-shot transferability of RL policies from NVIDIA Isaac Sim to real mobile robots with comparable performance to ROS Nav2, validating practical sim-to-real transfer on navigation and obstacle avoidance.
— ICLR 2025 workshop presentation on whole-body control transfer for arm-mounted quadrupeds, documenting that training dynamic behaviors from scratch often fails with pronounced sim-to-real gap on commodity hardware with complex actuators.
— Techman Robot (Quanta subsidiary) deployed NVIDIA Isaac Sim for electronics product inspection, achieving 20% cycle time reduction and 70% programming time savings through digital twin simulation and optimization.
— MIT RialTo pipeline builds digital twins from 3D phone scans for reinforcement learning, achieving 67% robustness improvement over imitation learning while acknowledging simulation exploitation challenges.
— MIT CSAIL research introduces LucidSim, combining generative AI with physics simulation for diverse robot training data. Robot dogs trained on LucidSim data achieved 88% parkour success vs 15% for expert-only baselines, outperforming domain randomization.
— NVIDIA ROSCon 2024 announcements highlight sim-first approach adoption, Isaac Sim partnerships with Universal Robots, Miso Robotics, and Wheel.me, signaling continued ecosystem maturity and vendor integration.
— NeurIPS 2024 research proposes using simulators to learn exploration policies when direct transfer fails, providing theoretical analysis and empirical validation showing polynomial sample complexity improvement over failed sim-to-real transfer.
— Meta-analysis of generative robot simulators reveals widespread challenge of low generalization capability across task generation methods, highlighting quality and robustness barriers despite methodological advances.
— ICRA 2024 Robotic Sim2Real Competition winner demonstrating consistent performance across simulation and real-world navigation/grasping/stacking with lightweight visual perception (11ms per frame) and sub-centimeter servo accuracy.
— ETH Zurich/Harvard/AI Center research develops learned residual physics framework improving soft robot simulation accuracy by 60%, reducing sim-to-real gap for deformable systems via differentiable FEM and neural network residuals.
— Market analysis forecasts robotic simulator market expansion of USD 1.89 billion at 23.3% CAGR (2023-2028), with APAC dominating 47% growth, signaling strong commercial adoption and ecosystem maturity.
— NEURA Robotics joins NVIDIA Humanoid Robot Developer Program using Isaac Sim/Lab for faster training; company targets delivery of up to 5 million humanoid and cognitive robots by 2030 using simulation-based learning.
— NVIDIA announces new simulation tools (NIM microservices, OSMO orchestration, teleoperation) at SIGGRAPH 2024 with Humanoid Robot Developer Program including Boston Dynamics, Fourier, Figure, 1X, and Skild AI.
— AutoMate framework (USC/NVIDIA) trains generalist assembly policy achieving 84.5% mean success rate in real-world zero-shot transfer across 20 different assemblies using RL+IL in simulation.
— NVIDIA Isaac Lab enables training Boston Dynamics Spot locomotion at 85,000-95,000 FPS over 4 hours with PPO and domain randomization, achieving zero-shot transfer to physical robot via Jetson Orin inference.
— Chiba University/collaborators publish Frontiers Bioengineering solution for data-imbalanced sim-to-real transfer, applied to soft exoskeleton finger joint stiffness estimation with 41-56% error reduction vs baselines.
— NVIDIA announces Isaac Perceptor (16.5M depth points/sec per camera, <1% translation error) and Isaac Manipulator with foundation models, adopted by ArcBest, BYD Electronics, KION, Teradyne Robotics, and others.
— NCSU/University of Michigan Nature publication demonstrates physics-informed sim-to-real exoskeleton training reducing metabolic energy by 24.3% walking, 13.1% running, 15.4% stair climbing in human subjects.
— IEEE Access paper on sim-to-real transfer for autonomous robot coverage path planning using egocentric frontier maps and semi-virtual environments, successfully transferring policies to real robot hardware.
— University of Alberta/Tokyo/NVIDIA AI Tech Centre publishes ICSE SEIP industrial benchmark for Isaac Sim manipulation across eight tasks, developing falsification framework for bridging simulation to physical robotics testing.
— Independent evaluation from Eurecat of NVIDIA Isaac Gym and Isaac Sim platforms for sim-to-real transfer, providing empirical assessment of platform fidelity and limitations on mobile manipulation tasks.
— CoRL 2024 open-source toolkit enabling scalable simulation-based evaluation of real-world robot manipulation policies (RT-1, Octo), supporting checkpoint selection and reproducibility across common robotic platforms.
— MIT/UW RialTo pipeline for robustifying imitation learning via RL in digital twin simulations constructed from real data, achieving 67% robustness improvement on diverse real-world manipulation tasks without extensive human data collection.
— Toyota's production deployment using NVIDIA Isaac Sim with READY ForgeOS for sim-to-real robotic programming in aluminum hot forging lines, eliminating safety risks and enabling seamless transfer of simulation-trained programs to live production cells.
— CoRL 2024 paper from Stanford/Google on human-in-the-loop sim-to-real framework for contact-rich manipulation (furniture assembly), outperforming domain randomization baselines and demonstrating scalable transfer for complex assembly tasks.
— RSS 2024 paper from UPenn/NVIDIA automating domain randomization and reward function design via LLMs, reducing manual tuning and enabling robust transfer on quadruped locomotion and dexterous manipulation tasks.
— ICAR 2023 paper on alternative approach: online DRL training directly on physical vehicle with safety supervisor, demonstrating improved sample efficiency and driving performance vs simulation training—challenging simulation necessity.
— CoRL 2023 conference paper introducing COMPASS method via differentiable causal discovery to close sim-to-real gap, achieving significant improvements in trajectory alignment and task success rate on challenging manipulation tasks.
— Hybrid offline-online framework for contact-rich manipulation (assembly, pivoting, screwing) using model-free RL with domain randomization and online force sensor adaptation, showing effective transfer on high-precision tasks.
— IEEE Transactions peer-reviewed survey categorizing sim-to-real transfer techniques (accurate simulators, kinematic/dynamic models, hierarchical controllers, demonstrations) confirming practical deployment pathway for RL on physical bioinspired robots.
— Research systematically exploring visual domain randomization for manipulation achieving 93% real-world success rate, demonstrating simulator-trained policies outperform limited real-world data baselines on diverse challenging tasks.
— NVIDIA DeXtreme project trained robot hand manipulation entirely in simulation using Isaac Gym and domain randomization, transferring successfully to real Allegro Hand with robust performance even under hardware faults.
— Trimble's real-world deployment with Boston Dynamics Spot for door detection using Isaac Sim and structured domain randomization, improving average precision from 5% to 87% in indoor environments.
— HaDR method using domain randomization to generate synthetic RGB-D datasets for hand localization in cluttered industrial environments, achieving improved performance over state-of-the-art datasets.
— Peer-reviewed empirical study on visual navigation challenging the assumption that higher simulation fidelity improves transfer, demonstrating that lower-fidelity simulators yield better generalization and faster training.
— Real-to-Sim-to-Real framework using 3D Gaussian Splatting for photorealistic simulation, enabling RGB-only zero-shot sim-to-real transfer for visual navigation in household and factory environments.
— Frontiers peer-reviewed research on visual non-prehensile manipulation via latent-space dynamics model, enabling transfer from simulation with reduced real-world data collection via feature extractor adaptation.
— Aalto University doctoral thesis synthesizing sim-to-real transfer methodologies including parameter estimation, meta-learning, and safety-aware adaptation, demonstrating successful policy transfer with dedicated algorithms.
— Peer-reviewed research examining reality gap in robot swarm simulation, documenting that discrepancies between simulation and real-world swarm behavior necessitate expensive real-world testing and limit predictive capability.
— Industry coverage documenting NVIDIA Isaac Sim and Omniverse adoption at scale by Amazon (warehouse design simulation and robot training) and PepsiCo (workflow optimization), signaling mature commercial ecosystem by mid-2022.
— Competition paper organized by Max Planck Institute and DeepMind on dexterous manipulation with pre-stage simulation training and real-robot stage on TriFinger platforms, highlighting sim-to-real transfer challenges and solutions.
— IEEE Transactions on Robotics paper on sim2real domain randomization for object detection achieving 86.32% mAP50 zero-shot and 97.38% one-shot transfer on real-world robotic perception tasks with synthetic dataset training.
— Google research demonstrating i-Sim2Real iterative sim-to-real method for human-robot interaction achieving 22 average and 150 maximum successive rally hits in real-world robotic table tennis deployment with industrial arm.
— UC Berkeley research demonstrating Dreamer world-model approach enabling direct real-world robot training without simulators—teaching quadrupeds to walk in one hour and training visual manipulation tasks, challenging reliance on simulation infrastructure.
— University of Michigan research on data augmentation for deformable object manipulation achieving 40% improvement in sim and near-doubling of physical success rates, demonstrating accelerated learning for soft object handling.
— Applied research from AIST Japan combining curriculum learning with domain randomization achieving 86% success on real-world precision insertion tasks (±0.01mm tolerance), demonstrating practical viability for high-precision manufacturing.
— Frontiers in Robotics and AI peer-reviewed survey synthesizing domain randomization techniques across TU Darmstadt, Honda Research, NVIDIA, and Google, documenting consolidated state-of-art in sim-to-real methodology.
— University of Tokyo research on ROBOTIS-OP3 humanoid successfully transferring bipedal locomotion policies to handle uneven terrain and severe disturbances without force/torque feedback, advancing actuator-constrained locomotion transfer.
— IEEE IROS 2022 paper from University of Toronto achieving 100% sim-to-real transfer success with surgical da Vinci endoscope through domain randomization, demonstrating effectiveness in high-fidelity medical imaging domain.
— Independent academic comparison of Gazebo and NVIDIA Isaac Sim for ROS-based mobile robotics simulation, providing peer-reviewed assessment of ecosystem maturity and platform capabilities.
— Comprehensive arXiv review paper on domain randomization techniques for robot learning, synthesizing methodological advances and identifying parameter selection and distribution design as critical bottlenecks in 2021.
— Google researcher practitioner analysis documenting extensive sim-to-real deployment across locomotion, navigation, and manipulation while highlighting necessity for offline evaluation and persistent hardware wear constraints.
— Meta AI's Habitat 2.0 platform launch in June 2021 achieving 1,200 SPS performance (850× faster than prior art) with ReplicaCAD dataset, demonstrating major vendor ecosystem investment in scalable simulation infrastructure.
— MIT research introducing PlasticineLab environment with physics-aware gradient-based planning for deformable object manipulation, extending sim-to-real training capabilities beyond rigid-body tasks presented at ICLR 2021.
— Applied research achieving 99.8% sim success but 67% real-world zero-shot transfer on soft continuum arm visual servoing, demonstrating both capability and performance degradation in sim-to-real transfer.
— IEEE Access peer-reviewed survey paper synthesizing sim-to-real transfer techniques, providing comprehensive academic analysis of the field's maturity and persistent methodological challenges in 2021.
— Research paper presenting D²-GMBC framework using dynamics and domain randomized offline RL for legged robots traversing uneven terrain, with real quadrupedal robot validation showing significant improvement over hand-tuned baselines.
— R:SS 2020 workshop paper providing critical assessment of sim-to-real limitations in agricultural robotics, documenting that current simulation techniques cannot achieve the accuracy and fidelity required for complex real-world tasks.
— ICRA 2020 research from NVIDIA demonstrating sim-to-real directional semantic grasping using domain randomization, with real-world validation showing promising results in bridging the reality gap for manipulation.
— RA-L/ICRA research proposing Bayesian Domain Randomization for adaptive parameter learning, reducing prior knowledge requirements while achieving effective sim-to-real transfer on robotic tasks.
— NVIDIA's GTC 2020 demonstration of Isaac Sim 2020.1 as the industry's first robotic AI development platform with integrated simulation, navigation, and manipulation tools, signaling vendor ecosystem maturity.
— RSS 2020 paper proposing task-oriented exploration framework for identifying system parameters in sim-to-real transfer, with real-world experiments on pouring and object dragging showing superior performance over task-agnostic methods.
— Critical research from Georgia Tech/Facebook AI Research showing agents exploit simulator imperfections, achieving low sim-to-real correlation (0.18) in embodied navigation—revealing methodological challenges in current evaluation.
— IROS 2020 research paper demonstrating SPOT framework achieving 100% success rates on block-stacking and row-making tasks with direct sim-to-real transfer, validating end-to-end learning for multi-step manipulation.
— ICRA 2019 quantitative evaluation of simulator accuracy across multiple platforms, documenting significant discrepancies between simulation and physical manipulation—highlighting persistent challenges in sim-to-real transfer.
— Open-source implementation of Active Domain Randomization method from CoRL 2019, replacing uniform sampling with active search for challenging environments—advancing reproducible research in domain randomization.
— Peer-reviewed research showing body parameter randomization in simulation improves transfer of CPG controllers to physical compliant quadruped robots, advancing domain randomization techniques for locomotion.
— IROS 2019 research demonstrating two-stage system identification and Projected Universal Policy approach successfully creating three functional biped locomotion controllers on Darwin OP2, advancing transfer of dynamic control.
— CoRL 2018 paper introducing SPOTA algorithm for robust policy optimization via domain randomization and model ensembles, enabling policies trained purely in randomized simulation to transfer to real systems without real-world data.
— Research paper demonstrating deep sets encoding combined with modular RL for object sorting tasks, with policies learned in minutes of simplified simulation and direct hardware deployment without retraining.
— NVIDIA Isaac SDK launched with early access in 2018, providing integrated simulation capabilities, perception libraries, and APIs—establishing a major commercial platform for simulation-based robot development.
— ICRA 2018 landmark paper from OpenAI/UC Berkeley demonstrating dynamics randomization enables robotic arm policies trained entirely in simulation to transfer directly to real hardware without loss of performance.
— Georgia Tech research paper presenting two-stage system identification approach successfully transferring three locomotion controllers to Darwin OP2 robot hardware in 25 trials, validating biped sim-to-real transfer.