Sim-to-Real Transfer

Sim-to-real transfer is getting a policy trained in simulation to work on the physical robot, through four families of methods: domain randomizationdomain randomizationDomain randomization is training a policy across a distribution of simulators with randomized visual appearance, dynamics or both, so the real world reads as one more sample. Tobin et al. randomized rendering and transferred an object detector to a real arm. Peng et al. randomized dynamics and transferred a pushing policy with no real-world training. The cost is conservatism: such a policy is optimal for no single instance., teacher-student distillationteacher-student distillationTeacher-student distillation is training two policies in sequence for partial observability. A teacher sees privileged simulator state such as exact terrain friction or object pose; a student imitates it using only deployment observations. RMA instead trained a base policy with a privileged encoder, then a separate state-action-history module predicting its latent extrinsics for asynchronous online inference. Lee and colleagues distilled a privileged ANYmal teacher into a proprioceptive student that hikes challenging terrain without ever seeing it., system identificationsystem identificationSystem identification is fitting a simulator's parameters to a real robot's measured dynamics, shrinking the sim-to-real gap by calibration. Hwangbo et al. found actuator error the dominant transfer obstacle on ANYmal and replaced its miscalibrated analytic model with a learned network from joint-command history to realized torque. It attacks the gap's dynamics component directly; domain randomization only averages over it. and real-to-sim reconstruction. Aljalbout and colleagues' survey covers all four, splitting methods that reduce the gap from those that overcome it^{Source 11}11 Aljalbout et al. 2025, The Reality Gap in Robotics: Challenges, Solutions, and Best Practices. Elie Aljalbout, Jiaxu Xing, Angel Romero et al., Annual Review of Control, Robotics, and Autonomous Systems 2026 (accepted), 2025.

Photo: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

Domain randomization: train on a distribution of simulators

Jakobi and colleagues found in 1995 that sim-to-real correspondence depended on how much noise the simulation used^{Source 13}13 Jakobi et al. 1995, Noise and the reality gap: The use of simulation in evolutionary robotics. Nick Jakobi, Phil Husbands, Inman Harvey, Advances in Artificial Life (ECAL 1995), Lecture Notes in Computer Science, pp. 704-720. Tobin and colleagues randomized textures, lighting and camera pose for tabletop object localization; their best network localized objects to about 1.5 cm over 480 webcam images of eight geometric objects, though most runs used ImageNet initialization (random initialization nearly matched it given enough simulated data), and their abstract frames the real world as "just another variation" once the simulator varies enough^{Source 1}1 Tobin et al. 2017, Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World. Josh Tobin, Rachel Fong, Alex Ray et al., IROS 2017.

Peng and colleagues randomized link mass, joint damping, friction, controller gains, action timing and observation noise, maximizing expected return over that dynamics distribution. A recurrent policy trained only in simulation pushed a puck with a seven-DoF Fetch arm, with start and target positions randomized within a 30 cm square^{Source 2}2 Peng et al. 2018, Sim-to-Real Transfer of Robotic Control with Dynamics Randomization. Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba et al., ICRA 2018.

Photo: Kevin Zakka et al. / arXiv. Licence: CC BY 4.0.

OpenAI's 2018 hand policy randomized friction coefficients and object appearance and transferred to the physical robot^{Source 14}14 OpenAI et al. 2018, Learning Dexterous In-Hand Manipulation. OpenAI, Marcin Andrychowicz, Bowen Baker et al., 2018. Its Rubik's-cube successor added a curriculumcurriculum learningCurriculum learning is training on tasks that start easy and grow harder as the agent improves. Rudin et al. adjusted each robot's terrain level by its progress. Their documented curriculum-trained policy for simulation and deployment used 4,096 robots, 98,304 transitions per batch and 1,500 updates in under twenty minutes on an i9-11900k CPU and RTX A6000 GPU. Their separate flat-terrain headline is under four minutes on one workstation GPU. version, automatic domain randomization, which widens ranges as the policy succeeds^{Source 3}3 OpenAI et al. 2019, Solving Rubik's Cube with a Robot Hand. OpenAI, Ilge Akkaya, Marcin Andrychowicz et al., 2019.

Isaac Lab, the GPU-parallel training framework, ships ADR as a configurable curriculum, but mass, friction, contact offsets and joint armature change through PhysX CPU APIs. Most parameters randomize at runtime, mesh scale and collider type only before simulation starts^{Source 4}4 NVIDIA et al. 2025, Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. NVIDIA, Mayank Mittal, Yunrong Guo et al., 2025. SPI-Active notes that poorly tuned randomization "can lead to overly conservative policies"^{Source 15}15 Sobanbabu et al. 2025, Sampling-Based System Identification with Active Exploration for Legged Robot Sim2Real Learning. Nikhil Sobanbabu, Guanqi He, Tairan He et al., 2025.

Teacher-student distillation: privilege in, proprioception out

Teacher-student training splits sensing from control: a teacher learns from privileged simulator state, then a student with deployable sensing imitates it. Learning by Cheating used this two-stage form for driving^{Source 17}17 Chen et al. 2019, Learning by Cheating. Dian Chen, Brady Zhou, Vladlen Koltun et al., CoRL 2019, after asymmetric actor-critic had given only the critic full state^{Source 18}18 Pinto et al. 2017, Asymmetric Actor Critic for Image-Based Robot Learning. Lerrel Pinto, Marcin Andrychowicz, Peter Welinder et al., arXiv preprint, 2017.

Lee and colleagues distilled a privileged RL teacher, which saw simulator-only terrain and contact information, into a temporal convolutional student via its actions and latent features. The student uses two seconds of proprioceptiveproprioceptionProprioception is sensing the body's own state (joint positions and velocities, orientation, contact), as opposed to exteroception of the outside world through cameras or lidar. It is fast, cheap and never occluded. At ETH Zurich, Lee et al.'s ANYmal controller hiked mud, snow, rubble and vegetation on proprioception alone. Miki et al. kept it as the trusted fallback when the terrain map disagrees with the body. history, plus the command and current state, and one controller per robot generation took ANYmal-B and ANYmal-C robots through mud, snow, rubble and vegetation without environment-specific tuning^{Source 5}5 Lee et al. 2020, Learning Quadrupedal Locomotion over Challenging Terrain. Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen et al., Science Robotics 5(47), 2020. Isaac Lab's DextrAH-RGB example distills a privileged KUKA and Allegro teacher into a stereo-RGB network^{Source 4}4 NVIDIA et al. 2025, Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. NVIDIA, Mayank Mittal, Yunrong Guo et al., 2025, and VIRAL's authors report that scaling to tens of GPUs, up to 64, made teacher and student training reliable for an RGB humanoid that ran up to 54 loco-manipulation cycles^{Source 19}19 He et al. 2025, VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation. Tairan He, Zi Wang, Haoru Xue et al., arXiv preprint, 2025.

RMA trains a base policy and an environment-factor encoder jointly with RL, then regresses an adaptation module on simulated state-action histories to predict the latent extrinsics. The module reads 50 state-action steps (0.5 seconds) and updates asynchronously at about 10 Hz on the A1, while the base policy runs at 100 Hz on the latest estimate, with no online gradient updates. The paper reports adaptation in fractions of a second without real-world fine-tuning, and failures after large perturbations or multiple leg obstructions; its latent encodes behavior-relevant conditions without guaranteeing each physical parameter^{Source 6}6 Kumar et al. 2021, RMA: Rapid Motor Adaptation for Legged Robots. Ashish Kumar, Zipeng Fu, Deepak Pathak et al., RSS 2021. UP-OSI had earlier fed dynamics parameters predicted from recent history into a universal policy^{Source 20}20 Yu et al. 2017, Preparing for the Unknown: Learning a Universal Policy with Online System Identification. Wenhao Yu, Jie Tan, C. Karen Liu et al., RSS 2017.

Isaac Lab calls the failure mode an information gap due to input mismatch: the student must reconstruct unobserved states from partial observations such as images, hardest under heavy camera occlusion^{Source 4}4 NVIDIA et al. 2025, Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. NVIDIA, Mayank Mittal, Yunrong Guo et al., 2025. The teacher-student panel illustrates this with authored terrain and noise, so its errors rise by construction.

System identification: measure the gap instead

Randomization averages over model error, and system identification measures it. Hwangbo and colleagues fitted an actuator network for ANYmal's series-elastic actuators to real excitation data, predicting torque from joint-position errors and velocities now and 0.01 and 0.02 seconds earlier; policies trained with ideal or analytical actuator models failed to take a single step without falling^{Source 7}7 Hwangbo et al. 2019, Learning agile and dynamic motor skills for legged robots. Jemin Hwangbo, Joonho Lee, Alexey Dosovitskiy et al., Science Robotics 4(26), 2019. Tan and colleagues paired an identified actuator model and simulated latency with randomization^{Source 21}21 Tan et al. 2018, Sim-to-Real: Learning Agile Locomotion For Quadruped Robots. Jie Tan, Tingnan Zhang, Erwin Coumans et al., arXiv preprint, 2018. SPI-Active samples parameters in parallel with Fisher-information-driven excitation and beats baselines by 42 to 63% in locomotion tasks^{Source 15}15 Sobanbabu et al. 2025, Sampling-Based System Identification with Active Exploration for Legged Robot Sim2Real Learning. Nikhil Sobanbabu, Guanqi He, Tairan He et al., 2025, and Miller and colleagues tune Spot simulation against hardware with CMA-ES^{Source 22}22 Miller et al. 2025, High-Performance Reinforcement Learning on Spot: Optimizing Simulation Parameters with Distributional Measures. AJ Miller, Fangzhou Yu, Michael Brauckmann et al., arXiv preprint, 2025.

ASAP pretrains phase-conditioned tracking policies in IsaacGym on retargetedretargetingRetargeting is mapping motion recorded on one body, usually a human's, onto a robot with different proportions and joint limits, so human demonstrations become references the robot can track. Raw retargeting ignores the robot's physical constraints and can produce morphologically infeasible references. H2O and ASAP therefore add reinforcement learning that restores balance and contact feasibility on the real body. human motions, runs them on a Unitree G1, and records trajectories with motion capture and onboard sensors. A second PPO policy learns a delta action added before the simulator step, the tracking policy is fine-tuned with the frozen delta model in simulation, and only the tracking policy is deployed. For the hardware study the delta covers only four ankle DoFs of the 23-DoF model^{Source 8}8 He et al. 2025, ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. Tairan He, Jiawei Gao, Wenli Xiao et al., RSS 2025.

Evaluation separates IsaacGym-to-IsaacSim and IsaacGym-to-Genesis tests, with open-loop replay and closed-loop tracking over 43 motion sequences, and ASAP does not win every metric: in the easy IsaacSim closed-loop group its mean global joint-position error is 106 mm against SysID's 105 mm. On hardware, compared only with Vanilla, the error falls from 61.2 to 50.2 mm on kicking and from 159 to 112 mm on LeBron James's "Silencer"^{Source 8}8 He et al. 2025, ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. Tairan He, Jiawei Gao, Wenli Xiao et al., RSS 2025. ASAP reports better tracking on the held-out "Silencer" motion without establishing coverage of arbitrary unseen skills, and its stated hardware limits are motor overheating and damage, motion-capture dependence and the data needed for a full 23-DoF delta model^{Source 8}8 He et al. 2025, ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. Tairan He, Jiawei Gao, Wenli Xiao et al., RSS 2025. Swift's champion-level drone racing likewise combined simulation RL with physical-world data^{Source 23}23 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023.

Real-to-sim: rebuild the scene, keep the physics

SplatSim replaces mesh rendering with Gaussian splats reconstructed from the deployment scene, while PyBullet supplies physics. Across four UR5 tasks with 40 trials each, diffusion policies trained on its simulated demonstrations reached 86.25% zero-shot real success, against 97.5% for real-data policies, with training augmentations included^{Source 9}9 Qureshi et al. 2024, SplatSim: Zero-Shot Sim2Real Transfer of RGB Manipulation Policies Using Gaussian Splatting. Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun et al., 2024.

The setup uses a Robotiq 2F-85 gripper and two RealSense D455 cameras, and splat preparation needs manual robot segmentation, CAD-derived link bounds and ICP alignment. The setup figure lists RGB plus end-effectorend effectorAn end effector is the last link of a robot arm plus its gripper, hand or tool, whose pose the arm exists to place. Kinematics is conventionally the map from joint angles to that pose. Many learned manipulation policies act in end-effector deltas, others in joint targets; across 13,000+ real rollouts, joint space favored control stability and task space favored generalization. pose as inputs, while the problem statement says the policy uses only RGB at test time^{Source 9}9 Qureshi et al. 2024, SplatSim: Zero-Shot Sim2Real Transfer of RGB Manipulation Policies Using Gaussian Splatting. Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun et al., 2024.

RoboGSim's August 2025 revision chains a Gaussian Reconstructor, Digital Twins Builder, Scene Composer and Interactive Engine, where mesh assets and measured layout connect it to Isaac Sim. Novel-pose rendering and trajectory replay are evaluated separately. Closed-loop policies act on splat renders while Isaac Sim handles kinematics and collisions. On UR5 ring-toss, ten trials with up to three grasps each, the real-data model placed 90% on the real robot but 30% in RoboGSim, so its scores cannot replace real-robot tests^{Source 10}10 Li et al. 2024, RoboGSim: A Real2Sim2Real Robotic Gaussian Splatting Simulator. Xinhai Li, Jialin Li, Ziheng Zhang et al., 2024.

Newer twins go beyond static appearance: RialTo runs RL in digital twins built from small amounts of real data, raising imitation-policy robustness by over 67%^{Source 24}24 Torne et al. 2024, Reconciling Reality through Simulation: A Real-to-Sim-to-Real Approach for Robust Manipulation. Marcel Torne, Anthony Simeonov, Zechu Li et al., 2024, and PolaRiS turns short video scans into environments that track real generalist-policy performance better than existing benchmarks^{Source 25}25 Jain et al. 2025, PolaRiS: Scalable Real-to-Sim Evaluations for Generalist Robot Policies. Arhan Jain, Mingtong Zhang, Kanav Arora et al., 2025. SimFoundry's simulated evaluations correlated with real results at a mean Pearson coefficient of 0.911 across 7 tasks and 5 policy architectures^{Source 31}31 Ranawaka et al. 2026, SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation. Nadun Ranawaka, Josiah Wong, Wei-Lin Pai et al., 2026, and NVIDIA's Newton 1.0 post describes a tiled camera sensor that ray-traces both triangle meshes and Gaussian splats^{Source 12}12 Reist et al. 2026, Newton Adds Contact-Rich Manipulation and Locomotion Capabilities for Industrial Robotics. Philipp Reist, Miguel Zamora Mora, JC Chang et al., 2026.

A rebuilt twin can still mislead. Kadian and colleagues measured a sim-versus-real success correlation of 0.18 in a 3D-scanned lab because agents exploited collision dynamics, and tuning simulation parameters raised it to 0.844^{Source 26}26 Kadian et al. 2019, Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance?. Abhishek Kadian, Joanne Truong, Aaron Gokaslan et al., IEEE Robotics and Automation Letters 2020, 2019.

Kaiyuan Chen, Ken Goldberg and colleagues at UC Berkeley and NVIDIA (October 2026) built RoboSimStudio, which lets people and coding agents edit the solver, contact model and physical parameters of a Newton simulation^{Source 32}32 Chen et al. 2026, RoboSimStudio: An Interactive Toolkit for Human and Agent Rapid Real2Sim of Robot Manipulation Tasks. Kaiyuan Chen, Shuangyu Xie, Uksang Yoo et al., 2026. In a pilot, coding agents built a bimanual sponge-wiping simulation from a smartphone video in about an hour, and a policy improved there completed the wipe on the real robot^{Source 32}32 Chen et al. 2026, RoboSimStudio: An Interactive Toolkit for Human and Agent Rapid Real2Sim of Robot Manipulation Tasks. Kaiyuan Chen, Shuangyu Xie, Uksang Yoo et al., 2026. One real rollout then displaced a bar support that the simulation held fixed, and the authors judged sim and real agreement on this single task mainly by qualitative observation^{Source 32}32 Chen et al. 2026, RoboSimStudio: An Interactive Toolkit for Human and Agent Rapid Real2Sim of Robot Manipulation Tasks. Kaiyuan Chen, Shuangyu Xie, Uksang Yoo et al., 2026.

Levers behind successful transfer

The survey says the action space "plays a crucial role in reducing the sim-to-real gap", citing joint-velocity control for manipulation, and that co-training with real-world data benefits robotic foundation models such as GR00T N1^{Source 11}11 Aljalbout et al. 2025, The Reality Gap in Robotics: Challenges, Solutions, and Best Practices. Elie Aljalbout, Jiaxu Xing, Angel Romero et al., Annual Review of Control, Robotics, and Autonomous Systems 2026 (accepted), 2025. Maddukuri and colleagues report that simulation data raises real-world performance by 38% on average^{Source 27}27 Maddukuri et al. 2025, Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation. Abhiram Maddukuri, Zhenyu Jiang, Lawrence Yunliang Chen et al., 2025, and Lei and colleagues identify structured representation alignment as its primary effect^{Source 28}28 Lei et al. 2026, A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies. Yu Lei, Minghuan Liu, Abhiram Maddukuri et al., arXiv preprint, 2026.

Deployed systems lean on such levers: Haarnoja and colleagues' soccer humanoid transferred through high-frequency control, targeted dynamics randomization and training perturbations^{Source 29}29 Haarnoja et al. 2023, Learning Agile Soccer Skills for a Bipedal Robot with Deep Reinforcement Learning. Tuomas Haarnoja, Ben Moran, Guy Lever et al., arXiv preprint (Science Robotics 2024), 2023. Actuator reality shaping makes physical actuators match the simulator's idealized dynamics^{Source 30}30 Yamamori et al. 2026, Actuator Reality Shaping for Zero-Shot Sim-to-Real Robot Learning. Satoshi Yamamori, Koji Ishihara, Kenjiro Minamikawa et al., arXiv preprint, 2026, and Boston Dynamics reports fewer fidelity issues because Atlas's model and hardware are extremely close^{Source 16}16 Rodriguez et al. 2026, Training a Humanoid Robot for Hard Work. Alberto Rodriguez, Shane Rozen-Levy, Vinay Kamidi, Boston Dynamics blog, 2026.

Where each family breaks

The two interactives construct their costs; measured costs come from the sources. ASAP's broader skill coverage and affordable real-robot data collection remain open^{Source 8}8 He et al. 2025, ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. Tairan He, Jiawei Gao, Wenli Xiao et al., RSS 2025.