Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies
Yu Lei, Minghuan Liu, Abhiram Maddukuri, Zhenyu Jiang, Yuke Zhu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine classifier-free guidance with adversarial domain alignment (CFG-ADDA, lambda=-0.5) | sim-and-real co-training · Real-world manipulation avg | success count: 15.3/30 -> 21/30 | Improved |
| set CFG-ADDA guidance scale lambda to -0.5 instead of 0 (ablation) | CFG-ADDA · Real-world manipulation avg | success count: 18.6/30 -> 21/30 | Improved |
| add optimal-transport regularization | co-training · Real-world manipulation avg | success count: 15.3/30 -> 14.3/30 | No clear effect |
| add adversarial domain adaptation (ADDA) | co-training · Real-world manipulation avg | success count: 15.3/30 -> 14.3/30 | No clear effect |
| add classifier-free guidance (lambda=0) | co-training · Real-world manipulation avg | success count: 15.3/30 -> 15.3/30 | No clear effect |
| add co-training with simulation data | real-only policy · Real-world manipulation avg | success count: 8.6/30 -> 15.3/30 | Improved |
| add OT/ADDA alignment methods under balanced mixing ratios | co-training · sim-and-sim visual-physics | success rate | Improved |
| add OT/ADDA alignment under unbalanced mixing ratios | co-training · sim-and-sim visual-physics | success rate | Hurt |
| add CFG under unbalanced mixing | co-training · sim-and-sim visual-physics | success rate | Improved |
| remove gradient reversal layer to promote domain-discriminative features (ablation) | co-training with ADDA · sim-and-sim vis-phys | success rate | Hurt |
| vary mixing ratio w in toy example with misaligned representations (ablation) | co-training toy diffusion model · toy L2 loss | L2 loss | No clear effect |
| structured representation alignment vs mixing ratio effect (ablation) | toy co-training · toy L2 loss ANOVA | variance explained: around 50% vs 20% | Improved |
Closest papers by meaning
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.90)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.88)
- Self-Improving Embodied Foundation Models (NeurIPS 2025 · similarity 0.88)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.87)
- Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
