Research map / Vision-language and multimodal / Robot action and navigation
Scaling up Memory for Robotic Control via Experience Retrieval
Ajay Sridhar, Jennifer Pan, Satvik Sharma, Chelsea Finn
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 keyframe memory (MemER) vs no history | high-level policy · Counting | wrong scoops: 61 -> 1 | Improved |
| add keyframe memory (MemER) vs 32-frame long history | high-level policy · Object Search optimal path | count of 60: 41 -> 57 | Improved |
| extend history from 8 to 32 frames | high-level policy · Counting | wrong scoops: 26 -> 12 | Improved |
| use MemER high-level policy vs Human HL oracle | high-level policy · Object Search / Counting | objects retrieved: 59 vs 58; 1 vs 0 wrong scoops | No clear effect |
| fine-tuned Qwen2.5-VL high-level policy instead of GPT-5 | high-level policy · Object Search offline | trajectory accuracy: 0.15 -> 0.80 | Improved |
| fine-tuned Qwen2.5-VL instead of Gemini Robotics-ER 1.5 | high-level policy · Counting offline | trajectory accuracy: 0.13 -> 0.67 | Improved |
| replace visual keyframes with text subtasks (ablation) | MemER · Object Search | optimal path used: 57 -> 28 | Hurt |
| add text subtasks to visual keyframe memory (ablation) | MemER · Counting | wrong scoops: 1 -> 13 | Hurt |
| multi-task instead of single-task training (ablation) | MemER · Object Search and Dust & Replace, original tasks | success | No clear effect |
| multi-task instead of single-task training (ablation) | MemER · Cross-task object generalization | success | Improved |
| merge fine-tuned high-level weights with pretrained Qwen2.5-VL (ablation) | MemER · three tasks | success | Improved |
Closest papers by meaning
- Remembering What Matters: From Markovian to Subtask-Causal Memory in VLA Policies (NeurIPS 2026 · similarity 0.93)
- ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning (NeurIPS 2026 · similarity 0.90)
- STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning (ICLR 2025 · similarity 0.89)
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.89)
- Learning with Language-Guided State Abstractions (ICLR 2024 · 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.
