← Back
ICRA 2026
Real-Time Generation of Near-Minimum-Energy Trajectories Via Constraint-Informed Residual Learning
Domenico Dona', Giovanni Franzese, Cosimo Della Santina, Paolo Boscariol, Basilio Lenzo
AI summary
A residual learning framework generates near minimum-energy trajectories in real-time, achieving up to 87.3% of optimal performance while running 100 to 1,000 times faster than traditional optimization methods.
Problem
Traditional minimum-energy trajectory planning relies on solving nonlinear optimal control problems, which are computationally expensive and cannot meet real-time requirements for adaptive robotic applications.
Approach
The method uses residual learning to predict only the correction needed to steer a standard trajectory toward optimality, embedding boundary conditions as hard constraints and leveraging probabilistic regressors for uncertainty-aware planning.
Key results
- Achieves 87.3% of optimal performance near training data and 50.8% far from it
- Reduces computation time by two to three orders of magnitude compared to OCP solvers
- Embeds boundary conditions as hard constraints while requiring less training data
- Integrates active learning and uncertainty quantification to guide data collection
Why it matters
Enables real-time, energy-efficient motion planning for industrial robots, directly supporting sustainable manufacturing and adaptive automation.
Abstract
No abstract on file.