Research map / Reinforcement learning / Visual and representation RL
Contrastive Modules with Temporal Attention for Multi-Task Reinforcement Learning
Siming Lan, Rui Zhang, Qi Yi, Jiaming Guo, Shaohui Peng, Yunkai Gao, Fan Wu, Ruizhi Chen, Zidong Du, Xing Hu, Xishan Zhang, Ling Li and 1 more
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 contrastive loss among modules (ablation) | CMTA · MT10-Mixed | max smoothed success rate: 79.46 -> 82.07 | Improved |
| add contrastive loss among modules (ablation) | CMTA · MT50-Mixed | max smoothed success rate: 62.66 -> 71.69 | Improved |
| add contrastive loss to CARE (ablation) | CARE · MT10-Mixed | max smoothed success rate: 61.35 -> 65.24 | Improved |
| add contrastive loss to CARE (ablation) | CARE · MT50-Mixed | max smoothed success rate: 45.00 -> 47.61 | Improved |
| add contrastive loss (ablation) | CMTA · MT10-Fixed and MT50-Fixed | success rate | No clear effect |
| remove temporal information (LSTM state) from attention (ablation) | CMTA · MT10-Fixed and MT10-Mixed | success rate | Improved |
| increase experts from 2 to 6 (ablation) | CMTA · MT10-Mixed (1.0M steps) | max smoothed success rate: 52.53 -> 78.49 | Improved |
| increase experts from 6 to 20 (ablation) | CMTA · MT10-Mixed (1.0M steps) | max smoothed success rate: 78.49 -> 75.01 | Hurt |
| CMTA vs CARE | CMTA · MT50-Mixed | max smoothed success rate: 45.00 -> 71.69 | Improved |
| CMTA vs CARE | CMTA · MT10-Fixed | max smoothed success rate: 68.03 -> 78.95 | Improved |
| CMTA vs single-task SAC upper bound | CMTA · MT10-Mixed | max smoothed success rate: 71.11 -> 82.07 | Improved |
| CMTA vs single-task SAC upper bound | CMTA · MT10-Fixed | max smoothed success rate: 64.33 -> 78.95 | Improved |
Closest papers by meaning
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.85)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · similarity 0.85)
- Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Prioritized Soft Q-Decomposition for Lexicographic Reinforcement Learning (ICLR 2024 · similarity 0.85)
- Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.85)
- Decoupling Meta-Reinforcement Learning with Gaussian Task Contexts and Skills (AAAI 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.
