Research map / Reinforcement learning / Policy gradient and actor-critic
Towards Bridging the Gap between Large-Scale Pretraining and Efficient Finetuning for Humanoid Control
Weidong Huang, Zhehan Li, Hangxin Liu, Biao Hou, Yao Su, Jingwen Zhang
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 |
|---|---|---|---|
| add large-scale SAC pretraining before model-based finetuning (ablation) | LIFT · Booster T1 finetuning, 1.5 m/s | tracking of target speed | Improved |
| add world-model pretraining (ablation) | LIFT · Booster T1 finetuning, 1.5 m/s | finetuning sample efficiency | Improved |
| replace physics-informed world model with MBPO ensemble (ablation) | LIFT · Booster T1 finetuning, 1.5 m/s | episode return | Hurt |
| raise UTD ratio from 1 to 5-10 in SAC pretraining (ablation) | SAC pretraining · T1LowDimJoystick | convergence speed | Improved |
| use larger entropy coefficient alpha in finetuning (ablation) | LIFT finetuning · Booster T1, 1.5 m/s | convergence stability | Hurt |
| use multi-step autoregressive loss horizon (2 or 4) instead of 1 (ablation) | world model training · Booster T1, 1.5 m/s | stable convergence to target velocity | Improved |
| swap SAC for PPO under deterministic data collection (ablation) | SSRL framework preliminary study · Booster T1 target 0.2 m/s | sample efficiency | Hurt |
| finetune pretrained policy vs PPO/FastTD3/SAC baselines | Booster T1 sim-to-sim finetuning · Brax velocity tracking | convergence | Improved |
Closest papers by meaning
- Efficient Test-Time Adaptation For Robot Policies (NeurIPS 2026 · similarity 0.88)
- Time-Efficient Reinforcement Learning with Stochastic Stateful Policies (ICLR 2024 · similarity 0.88)
- Accelerating Visual-Policy Learning through Parallel Differentiable Simulation (NeurIPS 2025 · similarity 0.88)
- Zero-Shot Whole-Body Humanoid Control via Behavioral Foundation Models (ICLR 2025 · similarity 0.88)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.87)
- ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning (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.
