Research map / Architectures and efficiency / RNNs and normalization
Differentiable Spatial Planning using Transformers
Devendra Singh Chaplot, Deepak Pathak, Jitendra Malik
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 |
|---|---|---|---|
| Use long-range self-attention for spatial value propagation instead of local convolutional propagation | SPT versus GPPN · In-distribution overall | planning accuracy: 92.31 -> 99.41 | Improved |
| Use long-range self-attention for spatial value propagation instead of local convolutional propagation | SPT versus GPPN · Out-of-distribution overall | planning accuracy: 86.95 -> 95.18 | Improved |
| Use long-range self-attention for spatial value propagation instead of local convolutional propagation | SPT versus GPPN · End-to-end mapping and planning overall | planning accuracy: 63.91 -> 82.29 | Improved |
| Train the mapper end-to-end with noisy distances (σ=1) rather than dense, perfect supervision | SPT end-to-end mapping and planning · Manipulation | planning accuracy: 98.42 -> 95.83 | Hurt |
| Train the mapper with only 5 trajectories rather than dense supervision | SPT end-to-end mapping and planning · Manipulation | planning accuracy: 98.42 -> 97.18 | Hurt |
| Combine noisy distances with supervision from only 5 trajectories | SPT end-to-end mapping and planning · Manipulation | planning accuracy: 98.42 -> 94.34 | Hurt |
| Use sparse supervision from 5 trajectories rather than dense supervision | SPT end-to-end mapping and planning · Manipulation | map accuracy: 98.96 -> 80.15 | Hurt |
| Use sparse supervision from 5 trajectories rather than dense supervision | SPT end-to-end mapping and planning · Manipulation | planning accuracy: 98.42 -> 97.18 | No clear effect |
| Use SPT rather than Dijkstra | Navigation planner · M=50 maps | runtime per map (ms): 371.05 -> 18.35 | Improved |
Closest papers by meaning
- [Re] Differentiable Spatial Planning using Transformers (NeurIPS 2022 · similarity 0.89)
- Learning Cognitive Maps from Transformer Representations for Efficient Planning in Partially Observed Environments (ICML 2024 · similarity 0.87)
- Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference (NeurIPS 2024 · similarity 0.85)
- Versatile Navigation Under Partial Observability via Value-guided Diffusion Policy (CVPR 2024 · similarity 0.84)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.84)
- MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale (AAAI 2025 · similarity 0.84)
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.
