NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 25·8 min read
Sparse Compositional Flow Matching by geometric assembly from motion primitives
TL;DR
Existing generative models for robot trajectories treat motion as dense, unstructured sequences, wasting data and requiring many samples. Motion-Primitive Dictionary Learning and Structural Sparse Flow Matching with Geometric Constraints compose reusable motion fragments directly in physical space while enforcing spatial continuity.
✦ Why It Matters
Engineers can build more sample-efficient robot motion generators by explicitly modeling compositional structure rather than treating trajectories as monolithic signals.
Key Takeaways
- Decompose robot trajectories into reusable motion primitives aligned with subtask boundaries for better generalization.
- Use learnable length masks and binary indicators so primitives are reused verbatim across different placements.
- Apply differentiable geometric loss to enforce spatial continuity and temporal contiguity between adjacent motion segments.
- Evaluate on Open X-Embodied and 3DMoTraj; expect 19-21% error reduction over dense generative baselines.
Full Summary
Embodied AI systems (robots, underwater vehicles, mobile robots) generate motion trajectories—sequences of executable movements. Current generative models treat these as dense, point-by-point signals without modeling underlying structure, leading to sample inefficiency.
This work proposes that trajectories naturally decompose into reusable motion primitives (recurring motion fragments) aligned with subtask boundaries. The framework introduces two coupled components: Motion-Primitive Dictionary Learning, which learns atoms with variable lengths and binary start indicators that serve as reusable primitives, and Structural Sparse Flow Matching with Geometric Constraints, which generates binary placement matrices using duration-aware tokenization and differentiable geometric loss to enforce spatial continuity and temporal contiguity between adjacent primitives.
Evaluated on Open X-Embodiment and 3DMoTraj datasets, the method achieved state-of-the-art accuracy with 19.2% ADE (average displacement error) improvement and 21.0% FDE (final displacement error) improvement over baselines, reducing the FDE/ADE ratio from 1.8 to 1.07.