CART: Context-Aware Terrain Adaptation using Temporal Sequence Selection for Legged Robots

A team of researchers from the University at Buffalo and the University of Texas at Austin has introduced CART (Context-Aware Terrain Adaptation using Temporal Sequence Selection), a novel high-level AI controller designed to make legged robots significantly more stable on complex, off-road terrain. The system addresses a critical flaw in current methods known as the 'Visual-Texture Paradox,' where a robot's visual expectations of a surface (like grass) don't match the actual physical feedback (like slippery mud). CART solves this by fusing two sensory streams in real-time: exteroception from onboard cameras and proprioception from joint sensors and inertial measurement units (IMUs).

This multimodal approach allows the robot to build a robust, context-aware understanding of the terrain. The team rigorously evaluated CART using an ANYmal-C robot in the IsaacSim simulator and a Boston Dynamics SPOT robot for real-world experiments across varied surfaces. The key performance metric was vibrational stability at the robot's base. The results were striking: CART achieved an average 5% higher success rate in simulation and boosted real-world stability by up to 45% compared to state-of-the-art baselines. Crucially, it delivered these dramatic stability improvements without increasing the time taken to complete locomotion tasks, making it both safer and more efficient.