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, 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 **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.
Figures & tables
Figure 2: Importance reweighting reshapes the learned action distribution via reweighting score functions during training time.
Figure 3: Visualization of controlled toy example. We co-train ∼ 30 and ∼ 3000 samples from target and source domain. Vertically in each column, the differences of prediction samples showcase the impact of representation alignment; horizontally in each row, the differences showcase the shift of importance reweighting effect controlled by mixing ratio w .
Figure 4: L2 loss with sweeping mixing ratio and delta z distance, and ANOVA variance decomposition. When representations overlap (red line) or are disjoint (blue line), co-trained models are insensitive to mixing ratios. The importance reweighting effect (blue bar) can only explain 20% of the performance variance.
Figure 5: Visualizations of the designed sim-and-sim and sim-and-real tasks. In the physics-only setting, we vary the physical parameters of objects, including mass, friction, and size, while keeping the appearance similar. Across all target environments, we tune the robot controller configurations to match those in the real world.
Figure 6: UMAP visualization of latent features. We visualize the features after the vision stem and after the encoder trunk fϕ . Red and blue dots represent the real and simulation features, respectively. We show the results for a specific mixing ratio in this figure. More details and results are provided in the Appendix D.2 .
Figure 7: Correlation between representation alignment and success rate. We group the results of different tasks in each setting and compute the Spearman and Pearson correlation coefficients after normalization. Blue, yellow, and green-series dots represent the same task in Fig. 8 . Gray dash line indicates the fitted curve.
Figure 8: Success rate of sim-and-sim co-training with different domain gaps. Each data point is computed over three policy checkpoints, and each checkpoint is evaluated for 200 trials. The best performance is consistently achieved in the range of (0.016,0.3) .
NutAssembly
MugCleanup
MugHang
Avg
co-training
0.85
0.44
0.27
0.52
+discrimination
0.79
0.415
0.215
0.47
Table 1: Performance degrades after adding penalty for representation alignment. The experiments are run with w=0.1 .
Figure 9: Average performance on sim-and-sim experiments. We can find that alignment-based methods (+OT/+ADDA) get improvement in the balanced mixing ratio group, while discernibility-based methods (+CFG) get improvement mainly in the unbalanced mixing ratio group. More detailed results are available in the Appendix. D.1
Method
Success Rate
Avg
NutAssembly
MugCleanup
MugHang
Real-only
11/30
8/30
6/30
8.6/30
Mixing Ratio w
0.016
0.1
0.3
0.016
0.1
0.3
0.016
0.1
0.3
-
Co-Training
17/30
11/30
16/30
16/30
9/30
7/30
8/30
13/30
7/30
15.3/30
+ OT
15/30
17/30
11/30
8/30
15/30
15/30
11/30
9/30
4/30
14.3/30
+ ADDA
13/30
13/30
15/30
6/30
14/30
11/30
10/30
14/30
7/30
14.3/30
Table 2: Real-world policy performance under different co-training strategies.
Figure 10: Performance of different co-training methods on MugCleanup with w=0.1 in sim-and-sim settings. The plots are similar for all tasks and balanced mixing ratios.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: N/M scaling
Figure 12: Empirical measurement about training sample overlapping using Bhattacharyya coefficient. We compute it in the same way as in Song et al. (2025) , but include the distances between observation features as we are modeling the conditional distribution.
Figure 13: Detailed results of sim-and-sim evaluations.
Figure 14: Balanced mixing ratios are robust to different policy architectures. Each data point is computed over three policy checkpoints, and each checkpoint is evaluated for 200 trials. The best performance is consistently achieved in the range of (0.016,0.3) .
Task
Pretrain+FT
Co-training
NutAssembly
0.77
0.925
MugCleanup
0.215
0.495
MugHang
0.225
0.26
Appendix
Table 3: Performance comparison with Pre-train+FT.
Figure 15: Experiments on different dataset sizes. Each data point is computed over three policy checkpoints, and each checkpoint is evaluated for 200 trials. The best performance is consistently achieved in the range of (wn,wq) , which supports our guideline below.
wn=N+MN.
Appendix
Algorithm 2 Guideline for Co-Training Mixing Ratio Selection
Figure 16: Existence of three regimes in co-training.
Figure 17: More UMAP visualizations of the observation representations.
Figure 18: More UMAP visualizations of the deeper layer representations.
Sim-and-real co-training is an effective way to scale robot learning by augmenting limited real-world demonstrations with simulation data. However, existing approaches typically treat simulators as fixed data sources, relying on uniform generation, offline retrieval, or manual parameter alignment by domain experts. We formulate Active Generative Simulation: optimizing a simulator's data-generation distribution to maximize downstream real-world performance. To solve this problem, we propose MetaGenSim, a meta-learning framework that trains an auxiliary rater network via meta-gradients to automatically discover high-utility data-generation distributions within a user-specified simulation parameter space, without manual parameter alignment. Across simulation and real-world manipulation tasks, MetaGenSim outperforms uniform generation and retrieval baselines, achieves a 150% relative improvement in real-world full-task success over uniform generation, and matches expert manual alignment. We further show that distributions curated with a lightweight proxy policy transfer to VLA fine-tuning, and that MetaGenSim can exploit partial alignment through atomic skill stitching and object-receptacle composition when exact sim-and-real task alignment is unavailable. Project website: https://meta-gensim.github.io
We introduce BEACON--Best-Effort Adaptation for Cross-Domain Co-Training--a theory-guided framework for training generative robot policies with abundant source demonstrations and limited target demonstrations. BEACON casts cross-domain co-training as a discrepancy-aware importance-reweighting problem, jointly learning a diffusion-based visuomotor policy and per-sample source weights through an objective inspired by the Best-Effort Adaptation (BEA) bound. To make this approach practical for high-dimensional sequence policies, we develop scalable instance-level discrepancy proxies, stochastic alternating updates for policy and weights, and a multi-source extension that balances heterogeneous source domains. Across sim-to-sim, sim-to-real, and multi-source manipulation settings, BEACON improves robustness and data efficiency over target-only, fixed-ratio co-training, and feature-alignment baselines. Without an explicit alignment objective, BEACON also brings source and target representations closer together as an implicit result of discrepancy-aware co-training.
Antong Zhang, Han Qi, Heng Yang
Department of Computer Science, Brown University · School of Engineering and Applied Sciences, Harvard University
Simulation demonstrations can supplement scarce real-world data for robot policy co-training. However, the value of using data curation to actively select these demonstrations for sim-to-real co-training remains underexplored. Existing curation methods also lack a unified criterion for measuring trajectory-level utility and set-level coverage from closed-loop target behavior. To address these gaps, we present the first systematic study of data curation for sim-to-real robot policy co-training and propose Trajectory-level Utility and set-level Coverage Optimization (TUCO). TUCO uses influence functions to trace how each source demonstration affects target-domain scoring rollouts. Our key insight is that these effects can be decomposed into an overall contribution to target return and variation across rollouts, providing a common closed-loop basis for measuring trajectory utility and set coverage. We further propose a performance-aligned subset optimizer that combines these measures in a unified curation objective to reduce redundancy and select complementary demonstrations. Extensive experiments on RoboMimic and OmniReset establish the value of active simulation data curation for sim-to-real policy co-training and show that TUCO achieves state-of-the-art performance across single-simulator, sim-to-sim, and sim-to-real settings.
Ning Zhu, Mengfei Zhao, Yikai Tang +5
Stanford University · Nanyang Technological University · AXIS Robotics +4