Unitree B1 climbs stairs without sensors using torque-driven RL

Traditional reinforcement learning for legged robots relies on position-based control, which requires state estimation (e.g., linear velocity) and is less adaptable to varied terrain. These frameworks are also typically tested on small, lightweight quadrupeds that lack the torque needed for high-complexity tasks. The new work by Dowdy and Vaz flips the script by using torque-driven RL on the heavyweight Unitree B1 robot, trained entirely in Nvidia Isaac Sim and Isaac Lab.

The torque-based policy enables the B1 to walk up and down stairs without any exteroceptive sensors (no cameras, lidar, or depth sensors) and without needing to estimate its own velocity. The system achieves linear speeds of 3.5 m/s and angular speeds of 1.5 rad/s, matching or exceeding position-based alternatives while being more robust to terrain changes. The paper, accepted at IEEE/SICE SII 2026, demonstrates that torque control can make heavy-duty quadrupeds viable for real-world industrial inspection, search-and-rescue, and logistics tasks.