Research map / Architectures and efficiency / RNNs and normalization
Gated Path Planning Networks
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 |
|---|---|---|---|
| replace VIN's max-pooling recurrent update with a gated LSTM update | path-planning networks · 15×15 2D mazes: NEWS, Moore, and Differential Drive | %Opt: 93.9 -> 99.2 (NEWS); 96.3 -> 98.8 (Moore); 98.4 -> 99.3 (Differential Drive) | Helped |
| use a larger convolution kernel (F=11) in GPPN (ablation) | GPPN · 15×15 2D NEWS, K=20 | %Opt: 85.5 -> 97.7 | Helped |
| increase VIN kernel size to F=11 (ablation) | VIN · 15×15 2D NEWS, K=20 | %Opt: 93.9 -> 85.5 | Hurt |
| increase GPPN iterations from K=5 to K=20 (ablation) | GPPN · 15×15 2D NEWS | %Opt: 93.0 -> 99.2 | Helped |
| increase training data from 10k to 25k examples (ablation) | GPPN · 15×15 2D NEWS | %Opt: 97.8 -> 99.2 | Helped |
| increase learning rate from 1e-3 to 1e-2 (ablation) | GPPN · 15×15 2D NEWS | %Opt: 99.0 -> 19.4 | Hurt |
| increase learning rate from 1e-3 to 1e-2 (ablation) | VIN · 15×15 2D NEWS | %Opt: 92.0 -> 2.9 | Hurt |
| use the single shared input gate instead of full input gating (ablation) | GPPN · 2D maze test metrics | test performance | No clear effect |
| remove the reward-function skip connection from the GPPN recurrent layers (ablation) | GPPN · 2D maze tasks | performance | No clear effect |
| predict position-specific convolution weights from the maze design (ablation) | Hyper-VIN · 15×15 2D NEWS | %Opt: Hyper-VIN 75.1; VIN 85.8; Value Iteration 94.2 | Hurt |
| use GPPN instead of VIN for planning from first-person RGB map images | 3D ViZDoom planner · 3D ViZDoom Differential Drive | %Opt: 96.9 -> 91.4 | Hurt |
| use GPPN instead of VIN for planning from first-person RGB map images | 3D ViZDoom planner · 3D ViZDoom NEWS and Moore | %Opt: 79.0 -> 94.1 (NEWS); 84.6 -> 94.5 (Moore) | Helped |
Closest papers by meaning
- Scaling Value Iteration Networks to 5000 Layers for Extreme Long-Term Planning (ICML 2025 · similarity 0.91)
- Value Iteration Networks (NeurIPS 2016 · similarity 0.90)
- Highway Value Iteration Networks (ICML 2024 · similarity 0.88)
- Integrating Symmetry into Differentiable Planning with Steerable Convolutions (ICLR 2023 · similarity 0.87)
- Affordances-Oriented Planning Using Foundation Models for Continuous Vision-Language Navigation (AAAI 2025 · similarity 0.84)
- Evaluating Vision-Language Models as Evaluators in Path Planning (CVPR 2025 · similarity 0.83)
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.