SPAR: Search-trained Push Action Ranker for Efficient Navigation Among Movable Obstacles
We introduce the Search-trained Push Action Ranker (SPAR), which learns to prioritize push primitives from simulated search outcomes.
ProofPaper ↗
Key points
- Navigation Among Movable Obstacles (NAMO) requires a robot to rearrange objects that block access to its goal.
- For a pushing robot, progress may depend on a setup push that enables a later opening without itself clearing a passage.
- Across successful runs on $300$ simulated two-push keyhole problems, SPAR Search reduces median simulator calls by $7.0\times$, leading to a $3.8\times$ reduction in planning time relative to Random Search.
- In real-robot experiments, SPAR Policy succeeds in $37/40$ trials without a physics-based digital twin, compared with $33/40$ for SPAR Search and $27/40$ for Random Search.
Sources (1)
- [1]SPAR: Search-trained Push Action Ranker for Efficient Navigation Among Movable ObstaclesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 03:25 PM
We introduce the Search-trained Push Action Ranker (SPAR), which learns to prioritize push primitives from simulated search outcomes.
Navigation Among Movable Obstacles (NAMO) requires a robot to rearrange objects that block access to its goal.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026Training-Free Continual Learning for Multimodal Forecasting via Prompt Evolution
- Oct 19, 2026LEAF: LUT-aware Efficient Attention on FPGAs
- Oct 19, 2026NVIDIA Cosmos3-H-Dreams: Bringing Clinical Surgery to Life Through Generative Simulation
- Oct 19, 2026Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation