Research map / Reinforcement learning / Visual and representation RL
Policy Transfer with Strategy Optimization
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 |
|---|---|---|---|
| search the target environment for the best strategy using CMA-ES rather than relying on a robust or adaptive policy | policy transfer · Hopper DART-to-MuJoCo, dim(μ)=5 or 10 | task performance | Helped |
| increase randomized dynamics from dim(μ)=2 to dim(μ)=5 (ablation) | SO-CMA · Hopper DART-to-MuJoCo | task performance | Helped |
| increase randomized dynamics from dim(μ)=5 to dim(μ)=10 (ablation) | SO-CMA · Hopper DART-to-MuJoCo | task performance | No clear effect |
| use strategy optimization with CMA-ES instead of Hist or UPOSI | policy transfer · Walker2d DART-to-MuJoCo with latency | task performance | Helped |
| use strategy optimization with CMA-ES instead of Robust or Hist | policy transfer · HalfCheetah DART-to-MuJoCo with delay | task performance | Helped |
| search for a strategy with CMA-ES rather than use the robust policy | Minitaur actuator-model transfer · quadruped inaccurate-to-accurate actuator model | task performance | Helped |
| search for a strategy with CMA-ES rather than use baseline policies | policy transfer · Hopper rigid-to-deformable-foot transfer | task performance | Helped |
| use Bayesian optimization instead of CMA-ES for strategy search (ablation) | strategy optimization · Hopper and Quadruped | task performance | No clear effect |
| use feed-forward neural networks rather than an LSTM to learn the dynamics model (ablation) | model-based strategy optimization · Hopper DART-to-MuJoCo | task performance | Hurt |
| use an LSTM model-based optimizer instead of CMA-ES (ablation) | strategy optimization · Hopper DART-to-MuJoCo | task performance | No clear effect |
| train with low-dimensional randomized dynamics rather than a higher-dimensional family of dynamics (ablation) | policy transfer methods · Hopper DART-to-MuJoCo with different ankle joint limits | task performance | Hurt |
Closest papers by meaning
- Flow-based Domain Randomization for Learning and Sequencing Robotic Skills (ICML 2025 · similarity 0.90)
- Domain Randomization via Entropy Maximization (ICLR 2024 · similarity 0.90)
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.89)
- Efficient Test-Time Adaptation For Robot Policies (NeurIPS 2026 · similarity 0.88)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.88)
- SPiDR: A Simple Approach for Zero-Shot Safety in Sim-to-Real Transfer (NeurIPS 2025 · 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.