AATMAAATMA
Menu
Back to News
Robotics

Torque-Driven RL Enhances Heavy Quadruped Robot Movement

A new torque-driven reinforcement learning framework enables heavyweight quadruped robots to navigate rough terrain and stairs without exteroceptive sensors.

A newly detailed reinforcement learning framework utilizes direct torque control to improve the locomotion of heavyweight quadruped robots like the Unitree B1. Validated in Nvidia's Isaac Sim, the policy enables the robot to reach speeds of 3.5 m/s and successfully navigate stairs without relying on exteroceptive sensors. The torque-driven approach eliminates the need for complex velocity state estimation, making the platform far more adaptable to rough terrain.

Reinforcement learning has steadily advanced the field of legged robotics, allowing machines to adapt to challenging and uneven terrain. However, the majority of current locomotion frameworks rely heavily on position-based control. This traditional approach limits the adaptability of the policy across different terrain types and forces reliance on complex state estimation techniques, such as constant linear velocity tracking.

Furthermore, these frameworks are typically tested on small, lightweight robotic dogs. These lighter units lack the hardware viability required for high-complexity industrial tasks. Addressing this gap, a newly published study details a highly effective torque control framework designed specifically for heavyweight, high-torque quadrupeds.

Removing State Estimation Barriers

The new reinforcement learning policy allows a heavy robotic agent to traverse rough environments and track desired linear velocities without requiring absolute knowledge of its own current speed. By controlling torque directly rather than dictating precise joint positions, the system allows the robot's physical dynamics to naturally absorb terrain irregularities.

Researchers validated the framework in simulation using Nvidia's Isaac Sim and Isaac Lab environments. Applying the policy to the heavyweight Unitree B1 quadruped, the robot achieved impressive sustained speeds of 3.5 meters per second and rotational speeds of 1.5 radians per second.

Blind Stair Navigation

The most impressive capability demonstrated by the new policy is its proprioceptive resilience. The simulated Unitree B1 successfully walked both up and down flights of stairs without the aid of any exteroceptive sensors like cameras or LiDAR.

By relying purely on torque feedback and internal joint states, the quadruped can function in environments where external sensors might be blinded by dust, smoke, or poor lighting. This structural shift toward torque-driven reinforcement learning dramatically expands the operational potential for heavy-duty autonomous robots in disaster recovery and industrial inspection.