Reproduce “ELMP: Efficient Learning for Motion Planning via Analytical Policy Gradients” 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
Yixiao Li, Tifanny Portela, Jordis Herrmann, René Zurbrügg, Marco Hutter
ELMP, a framework from ETH Zürich and ABB, combines behavior cloning pre-training with self-supervised Analytical Policy Gradient (APG) fine-tuning and tool-aware geometric conditioning for neural motion planning. This approach enables data-efficient adaptation to novel environments and variable robot kinematics, achieving an 84.8% average success rate and reducing expert data requirements by up to 60 times while maintaining millisecond-level inference latency.
Key insights
- Analytical Policy Gradients applied to kinematic motion planning can significantly improve policy performance, reducing environmental collisions and enhancing accuracy without requiring additional expert demonstrations.
- Explicitly encoding variable tool geometry with point clouds provides superior collision avoidance and generalization compared to bounding box abstractions, especially in cluttered or challenging scenarios.
- Self-supervised APG fine-tuning offers a highly data-efficient mechanism for adapting NMPs to novel environments, reducing data generation costs by two orders of magnitude compared to Behavior Cloning with new expert data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.