Differentiable Motion Manifold Primitives for Reactive Motion Generation under Kinodynamic Constraints
Yonghyeon Lee
AI summary
Problem
Traditional trajectory optimization is too slow for real-time reactive control, while existing manifold-based methods fail to satisfy strict kinodynamic constraints.
Approach
The method learns a continuous-time trajectory manifold from offline optimizations and fine-tunes a neural decoder to directly enforce kinodynamic constraints during generation.
Key results
- Sub-10ms planning speed for dynamic throwing
- 100% task success and constraint satisfaction with rejection sampling
- Significantly faster than traditional trajectory optimization
- Robust generalization to unseen task parameters
Why it matters
Provides a scalable pathway for real-time reactive control in high-dimensional robotic systems operating under strict physical limits.
Abstract
Real-time motion generation – which is essential for achieving reactive and adaptive behavior – under kino- dynamic constraints for high-dimensional systems is a crucial yet challenging problem. We address this with a two-step approach: offline learning of a lower-dimensional trajectory manifold of task-relevant, constraint-satisfying trajectories, fol- lowed by rapid online search within this manifold. Extending the discrete-time Motion Manifold Primitives (MMP) frame- work, we propose Differentiable Motion Manifold Primitives (DMMP), a novel neural network architecture that encodes and generates continuous-time, differentiable trajectories, trained using data collected offline through trajectory optimizations, with a strategy that ensures constraint satisfaction – absent in existing methods. Experiments on dynamic throwing with a 7-DoF robot arm demonstrate that DMMP outperforms prior methods in planning speed, task success, and constraint satisfaction. Project page: https://diffmmp.github.io/.