Research map / Vision-language and multimodal / Robot action and navigation
FindingDory: A Benchmark to Evaluate Memory in Embodied Agents
Karmesh Yadav, Yusuf Ali, Gunshi Gupta, Yarin Gal, Zsolt Kira
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 |
|---|---|---|---|
| supervised fine-tune the VLM on FindingDory training data | Qwen2.5-VL · FindingDory high-level tasks | HL success rate: average improvement of 25% across tasks | Improved |
| add FindingDory SFT data to Video-R1 SFT (ablation) | Video-R1 · VSI-Bench | MCQ accuracy: 33.33 -> 35.06 | Improved |
| add FindingDory SFT data to Video-R1 SFT (ablation) | Video-R1 · VSI-Bench | numerical accuracy: 32.67 -> 32.35 | No clear effect |
| add FindingDory data to GRPO training (ablation) | Video-R1 · VSI-Bench | MCQ accuracy: 33.57 -> 35.42 | Improved |
| add FindingDory data to GRPO training (ablation) | Video-R1 · VSI-Bench | numerical accuracy: 33.9 -> 34.82 | Improved |
| increase number of input video frames for frozen VLMs (ablation) | frozen VLMs · FindingDory high-level tasks | HL success rate | No clear effect |
| train SFT model with longer (subsampled) videos (ablation) | Qwen SFT · FindingDory high-level tasks | HL success rate | Improved |
| replace video input with textual memory summaries (ablation) | Qwen2.5-VL text agent · FindingDory high-level tasks | HL success rate | Hurt |
| decompose policy: add ImageNav low-level policy after high-level VLM (ablation) | Qwen2.5-VL-7B hierarchical agent · FindingDory | LL-SR | Hurt |
| use mapping-based instead of ImageNav low-level policy (ablation) | Qwen2.5-VL-7B hierarchical agent · FindingDory | LL-SR: only 25% relative drop in SR | Improved |
| subsample videos to 96 frames (ablation) | FindingDory tasks · FindingDory solvability | task solvability: approx 1% drop | Hurt |
Closest papers by meaning
- Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration (CVPR 2026 · similarity 0.91)
- EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.91)
- Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning (CVPR 2025 · similarity 0.90)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.90)
- Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.89)
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.
