Safe and Stable Neural Network Dynamical Systems for Robot Motion Planning
Allen Emmanuel Binny, Mahathi Anand, Hugo Tadashi Kussaba, Lingyun Chen, Shreenabh Agrawal, Fares Abu-Dakka, Abdalla Swikir
AI summary
Problem
Traditional learning-from-demonstration methods struggle to guarantee both stability and obstacle avoidance in complex environments, often relying on computationally heavy online adjustments or overly conservative polynomial parameterizations that limit motion complexity.
Approach
The framework jointly trains a neural network dynamical system with neural Lyapunov and barrier certificates, then applies split conformal prediction to provide probabilistic guarantees on their correctness across the entire workspace.
Key results
- Jointly learns neural dynamics and formal stability/safety certificates from demonstrations
- Provides probabilistic PAC guarantees on certificate validity via split conformal prediction
- Successfully replicates complex 2D and 3D motions from synthetic and kinesthetic robot data
- Matches or exceeds baseline performance while handling complex obstacle configurations
Why it matters
Enables reliable, safe robot motion planning from demonstrations without real-time optimization, advancing practical deployment in industrial automation and human-robot interaction.
Abstract
Learning safe and stable robot motions from demon- strations remains a challenge, especially in complex, nonlinear tasks involving dynamic, obstacle-rich environments. In this letter, we propose Safe and Stable Neural Network Dynamical Systems S2-NNDS, a learning-from-demonstration framework that simul- taneously learns expressive neural dynamical systems alongside neural Lyapunov stability and barrier safety certificates. Unlike traditional approaches with restrictive polynomial parameteriza- tions, S2-NNDS leverages neural networks to capture complex robot motions, providing probabilistic guarantees through split conformal prediction in learned certificates. Experimental results in various 2D and 3D datasets—including LASA handwriting and demonstrations recorded kinesthetically from the Franka Emika Panda robot—validate the effectiveness of S2-NNDS in learning robust, safe, and stable motions from potentially unsafe demon- strations.