A new quadruped study found that duty factor, the share of each gait cycle that a foot stays on the ground, predicted recovery from small disturbances more consistently than the label “walk” or “trot.” The pattern appeared across three control approaches and ended with a Unitree Go2 crossing a 15-centimeter-wide beam.
Walk and trot usually describe the order and timing of foot contacts. Each label can still hide large changes in speed, stance width and contact time, so the researchers asked which continuous setting better explains robustness when those variables are separated.
Holding gait labels to the same contact time
The team first optimized motions for a model of the Ghost Robotics Spirit 40 and estimated whether a tiny departure from the planned motion would shrink or grow as the robot continued. At a matched duty factor of 0.75, meaning each foot stayed down for three-quarters of its cycle, the statistical comparison covered 20 walking and 20 trotting trajectories. Gait label did not explain a significant difference in this recovery measure, while speed and stance width did.
A second comparison used 60 more trajectories spanning 12 duty factors. Among contact time, speed and stance width, contact time had the strongest association with the recovery measure: keeping a foot down longer was linked to smaller errors as motion continued. This is an internal comparison among the parameters tested, not a claim that duty factor explains every source of stability.
From simulation to a narrow beam
The researchers next trained separate walk and trot policies for a simulated Unitree Go2, then varied duty factor and stance width while the robot attempted to cross a three-metre beam under forces up to 25 newtons and torques up to 3 newton-metres. At a 0.15-metre stance width, the 0.5-duty-factor trot succeeded in 31.8% of trials, while walking policies at duty factors of 0.8, 0.85 and 0.9 exceeded 70%; every setting failed at the narrowest 0.05-metre width.
A selector network then learned to choose the shortest contact time that kept success near the best observed rate for a given condition. It selected more contact time as the stance narrowed. That rule still had limits: 35 of 168 operating conditions were excluded because all tested settings almost always failed, and even the included 0.10-metre-wide stances remained unreliable.
The hardware test used a 0.15-metre beam, three commanded speeds and five runs for each gait-and-speed pairing, for 45 trials in total. Walk and trot at a 0.75 duty factor completed all 30 runs, while the 0.5-duty-factor trot completed 12 of 15; the robot relied on carefully calibrated motion capture for world-position estimates.
Tuning foot contact time
The result suggests that a locomotion system can tune foot contact time as a continuous robustness control instead of treating a change from walk to trot as the main decision.
The evidence covers walking and trotting, controlled timing choices and a small hardware sample under laboratory motion capture. Bounding, pronking and cantering were not tested, and the study did not establish how more contact time trades against energy use or how well the pattern survives noisier onboard vision.
Sources
- Zhu et al., “Duty Factor Predicts Robust Constrained Quadrupedal Locomotion Across Gait Types”, arXiv:2609.22073v1, 18 September 2026.
- Full paper in HTML, including methods, trial counts and limitations. The lead image is Figure 1 by the paper’s authors, reused under the paper’s CC BY 4.0 license.
- Robomechanics Lab Quad-SDK repository, the open-source control framework used for the physical-robot investigation.