A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies

• URL: http://arxiv.org/abs/2604.13645v1
• Date: Wed, 15 Apr 2026 09:14:43 GMT
• Title: A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies
• Abstract summary: Co-training combines limited in-domain real-world data with abundant surrogate data such as simulation or cross-embodiment robot data.<n>Despite its empirical success, the mechanisms that determine when and why co-training is effective remain poorly understood.
• Score: 28.828770839369536
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Co-training, which combines limited in-domain real-world data with abundant surrogate data such as simulation or cross-embodiment robot data, is widely used for training generative robot policies. Despite its empirical success, the mechanisms that determine when and why co-training is effective remain poorly understood. We investigate the mechanism of sim-and-real co-training through theoretical analysis and empirical study, and identify two intrinsic effects governing performance. The first, \textbf{``structured representation alignment"}, reflects a balance between cross-domain representation alignment and domain discernibility, and plays a primary role in downstream performance. The second, the \textbf{``importance reweighting effect"}, arises from domain-dependent modulation of action weighting and operates at a secondary level. We validate these effects with controlled experiments on a toy model and extensive sim-and-sim and sim-and-real robot manipulation experiments. Our analysis offers a unified interpretation of recent co-training techniques and motivates a simple method that consistently improves upon prior approaches. More broadly, our aim is to examine the inner workings of co-training and to facilitate research in this direction.

Related papers

• RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies [20.899640428403377]
  Existing benchmarks exhibit significant domain overlap between training and evaluation.<n>We introduce RoboLab, a simulation benchmarking framework designed to address these challenges.<n>By providing granular metrics and a scalable toolset, RoboLab offers a scalable framework for evaluating the true generalization capabilities of task-generalist robotic policies.
  arXiv  Detail & Related papers  (2026-04-10T19:42:21Z)
• Few-Shot Neural Differentiable Simulator: Real-to-Sim Rigid-Contact Modeling [9.084754731100311]
  We propose a few-shot real-to-sim approach that combines analytical formulations with the representational capacity of graph neural network (GNN)-based models.<n>Using only a small amount of real-world data, our method calibrates analytical simulators to generate large-scale synthetic datasets that capture diverse contact interactions.<n> Experimental results demonstrate that our approach enables learning-based simulators to outperform differentiable baselines in replicating real-world trajectories.
  arXiv  Detail & Related papers  (2026-03-06T12:32:56Z)
• D-REX: Differentiable Real-to-Sim-to-Real Engine for Learning Dexterous Grasping [66.22412592525369]
  We introduce a real-to-sim-to-real engine that leverages the Gaussian Splat representations to build a differentiable engine.<n>We show that our engine achieves accurate and robust performance in mass identification across various object geometries and mass values.<n>Those optimized mass values facilitate force-aware policy learning, achieving superior and high performance in object grasping.
  arXiv  Detail & Related papers  (2026-03-01T15:32:04Z)
• Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training [21.855770200309674]
  We propose a unified sim-and-real co-training framework for learning generalizable manipulation policies.<n>We show it can leverage abundant simulation data to achieve up to a 30% improvement in the real-world success rate.
  arXiv  Detail & Related papers  (2025-09-23T04:32:53Z)
• Revealing the Challenges of Sim-to-Real Transfer in Model-Based Reinforcement Learning via Latent Space Modeling [31.74241286023207]
  Reinforcement learning (RL) is playing an increasingly important role in fields such as robotic control and autonomous driving.<n>The gap between simulation and the real environment remains a major obstacle to the practical deployment of RL.<n>We propose a latent space based approach to analyze the impact of simulation on real-world policy improvement.
  arXiv  Detail & Related papers  (2025-06-15T06:02:42Z)
• Robotic World Model: A Neural Network Simulator for Robust Policy Optimization in Robotics [50.191655141020505]
  This work advances model-based reinforcement learning by addressing the challenges of long-horizon prediction, error accumulation, and sim-to-real transfer.<n>By providing a scalable and robust framework, the introduced methods pave the way for adaptive and efficient robotic systems in real-world applications.
  arXiv  Detail & Related papers  (2025-01-17T10:39:09Z)
• Reinforcement Learning with Human Feedback for Realistic Traffic Simulation [53.85002640149283]
  Key element of effective simulation is the incorporation of realistic traffic models that align with human knowledge. This study identifies two main challenges: capturing the nuances of human preferences on realism and the unification of diverse traffic simulation models.
  arXiv  Detail & Related papers  (2023-09-01T19:29:53Z)
• Model-Based Reinforcement Learning with Multi-Task Offline Pretraining [59.82457030180094]
  We present a model-based RL method that learns to transfer potentially useful dynamics and action demonstrations from offline data to a novel task. The main idea is to use the world models not only as simulators for behavior learning but also as tools to measure the task relevance. We demonstrate the advantages of our approach compared with the state-of-the-art methods in Meta-World and DeepMind Control Suite.
  arXiv  Detail & Related papers  (2023-06-06T02:24:41Z)
• Attention-based Adversarial Appearance Learning of Augmented Pedestrians [49.25430012369125]
  We propose a method to synthesize realistic data for the pedestrian recognition task. Our approach utilizes an attention mechanism driven by an adversarial loss to learn domain discrepancies. Our experiments confirm that the proposed adaptation method is robust to such discrepancies and reveals both visual realism and semantic consistency.
  arXiv  Detail & Related papers  (2021-07-06T15:27:00Z)
• Domain Adaptive Robotic Gesture Recognition with Unsupervised Kinematic-Visual Data Alignment [60.31418655784291]
  We propose a novel unsupervised domain adaptation framework which can simultaneously transfer multi-modality knowledge, i.e., both kinematic and visual data, from simulator to real robot. It remedies the domain gap with enhanced transferable features by using temporal cues in videos, and inherent correlations in multi-modal towards recognizing gesture. Results show that our approach recovers the performance with great improvement gains, up to 12.91% in ACC and 20.16% in F1score without using any annotations in real robot.
  arXiv  Detail & Related papers  (2021-03-06T09:10:03Z)
• Point Cloud Based Reinforcement Learning for Sim-to-Real and Partial Observability in Visual Navigation [62.22058066456076]
  Reinforcement Learning (RL) represents powerful tools to solve complex robotic tasks. RL does not work directly in the real-world, which is known as the sim-to-real transfer problem. We propose a method that learns on an observation space constructed by point clouds and environment randomization.
  arXiv  Detail & Related papers  (2020-07-27T17:46:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.