Research map / Reinforcement learning / Visual and representation RL
Reinforcement Learning with Action-Free Pre-Training from Videos
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 APV instead of DreamerV2 | DreamerV2 · Meta-world six manipulation tasks | aggregate success rate: 67.9% -> 95.4% | Helped |
| Use the stacked latent prediction model without pre-training or intrinsic bonus (ablation) | DreamerV2 · Meta-world manipulation tasks | success rate | No clear effect |
| Add generative pre-training on action-free videos (ablation) | APV · Meta-world manipulation tasks | success rate | Helped |
| Add the video-based intrinsic bonus (ablation) | APV · Meta-world manipulation tasks | success rate | Helped |
| Combine generative pre-training with the video-based intrinsic bonus (ablation) | APV · Meta-world manipulation tasks | success rate | Helped |
| Use naive fine-tuning that initializes the action-conditional model from the action-free model (ablation) | DreamerV2 · Meta-world manipulation tasks | success rate | No clear effect |
| Use APV's stacked fine-tuning architecture instead of naive fine-tuning (ablation) | pre-trained action-free model · Meta-world manipulation tasks | success rate: >10% higher success rate | Helped |
| Set intrinsic-bonus trajectory length to τ=5 instead of τ∈{1,3} (ablation) | APV · Meta-world manipulation tasks | success rate | Helped |
| Set intrinsic-bonus trajectory length to τ=10 instead of τ=5 (ablation) | APV · Meta-world manipulation tasks | success rate | Hurt |
| Transfer only the convolutional image encoder and decoder, omitting pre-trained dynamics (ablation) | APV · Meta-world manipulation tasks | success rate | Hurt |
| Add in-domain Meta-world videos to RLBench pre-training (ablation) | APV · Meta-world unseen manipulation tasks | success rate | No clear effect |
| Add in-domain Meta-world videos while freezing the pre-trained representation model (ablation) | APV · Meta-world unseen manipulation tasks | success rate | Helped |
Closest papers by meaning
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.90)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.89)
- The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training (NeurIPS 2024 · similarity 0.89)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
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.