Research map / Reinforcement learning / MDPs and planning
Understanding Domain Randomization for Sim-to-real Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use domain randomization with history-dependent policies (memory) | Sim-to-real transfer over a finite simulator class · Finite simulator class without separation condition | sim-to-real gap: from the order of H to the order of \sqrt{H} | Helped |
| Remove memory from the policy | Domain randomization · Finite simulator class without separation condition | sim-to-real gap: H | Hurt |
| Remove randomization across environments | Domain randomization · Finite simulator class | sim-to-real gap: H | Hurt |
| Require the simulator class to be smooth near the real MDP | Domain randomization oracle · Infinite simulator class | sim-to-real gap: O\left(\frac{D\sqrt{d_eH\log \left(H \cdot \mathcal{N}(\mathcal{F}, 1 / H)\right)}}{ \nu\left(\caC_{\caM^*,\epsilon}\rig | Helped |
| Use domain randomization without smoothness in a large simulator class | Domain randomization oracle · Large non-smooth simulator class with linear mixture models | sim-to-real gap: \Omega(H) | Hurt |
| Use adversarial training instead of optimizing average value over randomized MDPs | Sim-to-real transfer · Simulator class with general function approximation | sim-to-real gap: O\left(D\sqrt{d_e H \log \left(H \cdot \mathcal{N}(\mathcal{F}, 1/H)\right)}\right) | Helped |
| Assume communicating MDPs with bounded diameter | Domain randomization oracle · Sim-to-real transfer | sim-to-real gap: Without Assumption [ref], there exists a hard instance ... Gap(pidr) = Omega(H). | Helped |
| Identify the real MDP using informative state-action samples, then execute its optimal policy | Base policy for domain randomization with separated finite simulator class · Finite simulator class with separation condition | sub-optimality gap: O\left(\frac{DM^2 \log(MH) \log^2(SMH/\delta)}{\delta^4}\right) | Helped |
| Use optimistic exploration and elimination without requiring separated MDPs | Base policy for finite simulator class · Finite simulator class without separation condition | regret: O\left(D\sqrt{MH \log(MH)}\right) | Helped |
| Limit policy updates using an importance-score threshold | General Optimistic Algorithm · Infinite-horizon average-reward MDPs with general function approximation | regret: O\left(D\sqrt{d_e H \log \left(H \cdot \mathcal{N}(\mathcal{F}, 1/H)\right)}\right) | Helped |
Closest papers by meaning
- Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL (NeurIPS 2024 · similarity 0.90)
- SPiDR: A Simple Approach for Zero-Shot Safety in Sim-to-Real Transfer (NeurIPS 2025 · similarity 0.90)
- Domain Randomization via Entropy Maximization (ICLR 2024 · similarity 0.89)
- Flow-based Domain Randomization for Learning and Sequencing Robotic Skills (ICML 2025 · similarity 0.87)
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.85)
- EASI: Evolutionary Adversarial Simulator Identification for Sim-to-Real Transfer (NeurIPS 2024 · similarity 0.85)
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.