Research map / Reinforcement learning / Visual and representation RL
Policy Adaptation From Foundation Model Feedback
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 |
|---|---|---|---|
| Add play-and-relabel policy adaptation using foundation-model feedback | MDETR-encoder policy · put-shapes-in-bowls, compositional generalization | success rate: 17.8% to 35.0% | Helped |
| Add play-and-relabel policy adaptation using foundation-model feedback | MDETR-encoder policy · pack-unseen-objects | success rate: 48.4% to 63.8% | Helped |
| Add play-and-relabel policy adaptation using foundation-model feedback | MDETR-encoder policy · CALVIN unseen environment, five-instruction chains | success rate at chain step 5: 5% to 11% | Helped |
| Add play-and-relabel policy adaptation using foundation-model feedback | MDETR-encoder policy · sim-to-real transfer, four tasks | success rate: +49.6% | Helped |
| Use 40 rather than 10 or 20 play demonstrations per scene (ablation) | PAFF · put-shapes-in-bowls, compositional generalization | success rate: 36.0% -> 51.0%; 37.6% -> 51.0% | Helped |
| Increase play demonstrations from 10 to 20 per scene (ablation) | PAFF · pack-unseen-objects, protocol 1 | success rate: 69.7% -> 73.2% | Helped |
| Increase play demonstrations from 20 to 40 per scene (ablation) | PAFF · pack-unseen-objects, protocol 1 | success rate: 73.2% -> 72.8% | Hurt |
| Use ST-Adapter rather than 2D joint-attention for temporal reasoning (ablation) | PAFF relabeling model · CALVIN unseen environment | average successful sequence length: 1.60 -> 1.73 | Helped |
| Use ST-Adapter rather than 3D joint-attention for temporal reasoning (ablation) | PAFF relabeling model · CALVIN unseen environment | average successful sequence length: 1.53 -> 1.73 | Helped |
| Use ST-Adapter rather than divided-attention for temporal reasoning (ablation) | PAFF relabeling model · CALVIN unseen environment | average successful sequence length: 1.46 -> 1.73 | Helped |
| Replace CLIP with MDETR as the visual and language policy encoder (ablation) | Policy without play and relabel · put-shapes-in-bowls, compositional generalization | success rate: 28.0% -> 33.8% | Helped |
| Use data augmentation during policy training | MDETR-encoder policy · put-shapes-in-bowls, compositional generalization | success rate: 33.8% -> 34.4% | Helped |
Closest papers by meaning
- Self-Improving Embodied Foundation Models (NeurIPS 2025 · similarity 0.88)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.87)
- Solving New Tasks by Adapting Internet Video Knowledge (ICLR 2025 · similarity 0.87)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.87)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- Active Fine-Tuning of Multi-Task Policies (ICML 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.