Learning Neural Observer-Predictor Models for Limb-Level Sampling-Based Locomotion Planning
Abhijeet Mangesh Kulkarni, Ioannis Poulakakis, Guoquan (Paul) Huang
AI summary
Problem
Simplified kinematic models fail to capture the complex closed-loop dynamics of legged robots, leading to inaccurate predictions and unsafe planning in cluttered environments, while existing data-driven methods often lack full-body prediction or require environment-specific retraining.
Approach
The authors design a decoupled neural architecture that uses a stability-guaranteed observer to estimate latent states from proprioceptive history, which initializes a GRU-based predictor to rapidly generate thousands of parallel full-body trajectory rollouts for sampling-based planners.
Key results
- Provable Uniformly Ultimately Bounded stability guarantees for the neural observer
- GRU predictor enabling 1000 parallel trajectory rollouts in 13ms
- Seamless integration with MPPI planner for limb-level collision checking
- Hardware validation on a Vision60 quadruped demonstrating accurate prediction and safe navigation in narrow passages
Why it matters
Provides a robust, real-time foundation for safe autonomous navigation of dynamic legged robots in complex environments without relying on overly conservative kinematic approximations.
Abstract
Accurate full-body motion prediction is essential for the safe, autonomous navigation of legged robots, enabling critical capabilities like limb-level collision checking in cluttered environments. Simplified kinematic models often fail to capture the complex, closed-loop dynamics of the robot and its low-level controller, limiting their predictions to simple planar motion. To address this, we present a learning-based observer-predictor framework that accurately predicts this motion. Our method features a neural observer with provable Uniformly Ultimately Bounded (UUB) guarantees that provides a reliable latent state estimate from a history of proprioceptive measurements. This stable estimate initializes a computationally efficient predic- tor, designed for the rapid, parallel evaluation of thousands of potential trajectories required by modern sampling-based planners. We validated the system by integrating our neural predictor into an Model Predictive Path Integral (MPPI)- based planner on a Vision60 quadruped. Hardware experi- ments successfully demonstrated effective, limb-aware motion planning in a challenging, narrow passage and over small objects, highlighting our system’s ability to provide a robust foundation for high-performance, collision-aware planning on dynamic robotic platforms.