Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Habitat-Web: Learning Embodied Object-Search Strategies From Human Demonstrations at Scale
Check 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 |
|---|---|---|---|
| train by behavior cloning on human demonstrations rather than shortest-path demonstrations | ObjectNav policy · ObjectNav MP3D val | success: 4.4% -> 31.6% | Helped |
| train by behavior cloning on human demonstrations rather than shortest-path demonstrations | Pick&Place policy · Pick&Place new initializations val | success: 1.9 -> 17.6 { ±0.8} | Helped |
| add SemSeg and SGE to RGB and depth inputs (ablation) | ObjectNav policy · ObjectNav MP3D val | success: 22.7% -> 31.6% | Helped |
| remove all RGBD and semantic visual inputs (ablation) | ObjectNav policy · ObjectNav MP3D val | success: 31.6% -> 0.0% | Hurt |
| add 10k THDA demonstrations with exhaustive-search behavior (ablation) | ObjectNav policy · ObjectNav MP3D val | success: 32.4% -> 33.2% | Helped |
| add 10k Gibson demonstrations to MP3D demonstrations (ablation) | ObjectNav policy · ObjectNav MP3D val | success: 33.9%; 1.5% worse than MP3D-only | Hurt |
| increase the number of human demonstrations (ablation) | ObjectNav policy · ObjectNav MP3D val | success | Helped |
| increase the number of human demonstrations (ablation) | Pick&Place policy · Pick&Place validation | success | Helped |
| use inflection weighting instead of vanilla behavior cloning loss (ablation) | behavior-cloned policy · ObjectNav | performance | Helped |
| use shortest-path demonstrations instead of human demonstrations | Pick&Place policy · Pick&Place new environments test | success: 8.3 -> 0.2 | Hurt |
| generalize from seen object-receptacle initializations to new object-receptacle pairs | Pick&Place policy trained on human demos · Pick&Place new instructions | success: 17.5% -> 15.1% | Hurt |
| evaluate on new environments rather than new instructions | Pick&Place policy trained on human demos · Pick&Place new environments test | success: 15.1% -> 8.3% | Hurt |
Closest papers by meaning
- Synatra: Turning Indirect Knowledge into Direct Demonstrations for Digital Agents at Scale (NeurIPS 2024 · similarity 0.87)
- Habitat 3.0: A Co-Habitat for Humans, Avatars, and Robots (ICLR 2024 · similarity 0.87)
- Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- Reward-free World Models for Online Imitation Learning (ICML 2025 · similarity 0.87)
- Exploration-Driven Generative Interactive Environments (CVPR 2025 · similarity 0.86)
- A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search (NeurIPS 2025 · 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.