UC Berkeley's Go2-W robot races 8.7% faster with active roll control
A wheeled quadruped uses MPC and RL to bank into turns like a motorcycle
A team led by Marla Eisman at UC Berkeley's Model Predictive Control Lab has published a paper detailing how they made a wheeled quadruped robot (the Unitree Go2-W) race faster by actively controlling its body roll during high-speed cornering. The approach combines offline time-optimal raceline generation, an online MPC planner that minimizes the lateral Load Transfer Ratio (LTR), and a low-level whole-body RL policy deployed across all 16 actuators on the robot. The key innovation is using the robot's leg joints as an active suspension system—knee joints generate anti-roll torque to bank the robot into turns, similar to how a motorcycle leans.
The physical track experiments showed dramatic improvements over a baseline non-tilting controller: mean LTR dropped by 44%, fastest lap time improved by 8.7%, and peak lateral acceleration capability increased by 21.3% to 1.98 m/s². The robot maintained robust high-speed stability beyond the range of the baseline. The paper has been accepted to AVEC 2026 and includes supplementary code and video. This work demonstrates how combining model-predictive control with learned policies can push the limits of agile robotics.
- MPC and RL framework reduces lateral Load Transfer Ratio (LTR) by 44% on the Unitree Go2-W
- Lap time improved 8.7% and peak lateral acceleration increased 21.3% to 1.98 m/s²
- Leg joints act as active suspension, generating anti-roll torque to bank into turns
Why It Matters
Active roll control could transform high-speed legged robots for search-and-rescue, inspection, and autonomous racing.