Reproduce “Motion Planning Transformers: A Motion Planning Framework for Mobile Robots” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jacob J. Johnson, Uday S. Kalra, Ankit Bhatia, Linjun Li, Ahmed H. Qureshi, Michael C. Yip
Motion Planning Transformers (MPT) introduces a framework utilizing transformer networks to efficiently guide sampling-based motion planners by predicting promising search regions. The approach reduces planning time by 7-25 times and enables robust generalization to varying map sizes through a novel random shifting of position encoding during training, demonstrated for both holonomic and non-holonomic robots.
Key insights
- Transformer networks can effectively learn complex, long-range spatial dependencies in environment representations, making them suitable for global context understanding in motion planning.
- A novel training routine involving random shifting of position encodings enables transformer models to generalize robustly to map sizes significantly larger than those seen during training, overcoming a major limitation of prior learning-based planners.
- Learning a probabilistic mask of 'promising regions' and associated orientations allows efficient guidance for sampling-based planners, even for complex non-holonomic robotic systems, significantly reducing search space and planning time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.