Quadruped-control preprint proposes sensor-light adaptation

Researchers reported CARO, a preprint framework that adds a torque-level residual observation to a quadruped robot's reinforcement-learning control loop. The approach embeds a fixed-base Euler-Lagrange model and uses a disturbance observer to signal dynamics mismatch, without torque sensors, contact estimation or vision-based floating-base measurements. The authors report improved zero-shot robustness in simulation and sim-to-real tasks involving out-of-distribution payloads, center-of-mass shifts, terrain changes, abrupt dynamics changes and elevated-platform landings.