Research map / Reinforcement learning / Offline RL
StARformer: Transformer With State-Action-Reward Representations for Robot Learning
Jinghuan Shang, Xiang Li, Kumara Kahatapitiya, Yu‐Cheol Lee, Michael S. Ryoo
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 StAR-representations via Step Transformer | DT · Atari Breakout (offline RL) | episodic return: 76.9 -> 124.1 | Improved |
| add StAR-representations via Step Transformer | DT · DMC Cheetah (offline RL) | episodic return: 27.9 -> 275.3 | Improved |
| add StAR-representations via Step Transformer | DT · Atari Seaquest (offline RL) | episodic return: 1131 -> 781 | Hurt |
| add StAR-representations via Step Transformer | DT · Atari Boxing (offline RL) | episodic return: 78.3 -> 81.2 | No clear effect |
| add StAR-representations via Step Transformer | DT · Atari Breakout (imitation) | episodic return: 54.3 -> 103.1 | Improved |
| add StAR-representations via Step Transformer | DT · DMC Cheetah (imitation) | episodic return: 0.5 -> 231.9 | Improved |
| use patch embeddings in Step Transformer instead of conv (ablation) | StARformer · Atari Breakout (offline RL) | episodic return: 55.6 -> 124.1 | Improved |
| use conv features for pure state tokens instead of patches (ablation) | StARformer · Atari Breakout (offline RL) | episodic return: 38.0 -> 124.1 | Improved |
| add pure conv state tokens in Sequence Transformer (ablation) | StARformer · Atari Breakout (offline RL) | episodic return: 38.7 -> 124.1 | Improved |
| order fine-grained patches in Step then conv in Sequence (P+C vs C+P) (ablation) | StARformer · Atari Breakout (offline RL) | episodic return: 46.3 -> 124.1 | Improved |
| replace conv features with ViT patch embeddings in DT (ablation) | DT · Atari Breakout (offline RL) | episodic return: 70.7 -> 47.3 | Hurt |
| layer-wise attention over all tokens instead of summation (StAR Fusion) (ablation) | StARformer · Atari Breakout (offline RL) | episodic return: 29.2 -> 124.1 | Improved |
Closest papers by meaning
- SiT: Symmetry-invariant Transformers for Generalisation in Reinforcement Learning (ICML 2024 · similarity 0.87)
- Decision SpikeFormer: Spike-Driven Transformer for Decision Making (CVPR 2025 · similarity 0.87)
- STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization (ICML 2025 · similarity 0.87)
- Decision ConvFormer: Local Filtering in MetaFormer is Sufficient for Decision Making (ICLR 2024 · similarity 0.86)
- QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (NeurIPS 2024 · similarity 0.86)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.86)
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.
