Energy-Efficient Gait Algorithm Helps Bipedal Robots Walk Smarter
A groundbreaking new algorithm developed by researchers in China could redefine how bipedal robots move, offering a smarter, more energy-conscious approach to robotic locomotion. The study, led by Zhiqiang Lu of Xi’an University of Science and Technology and Henan University, introduces a novel gait-planning framework that significantly reduces energy consumption while maintaining high stability—two critical challenges in the development of humanoid robots.
As robotics continues to advance, the demand for machines that can operate autonomously in complex, human-centric environments grows. Whether for industrial inspection, disaster response, or personal assistance, bipedal robots must walk efficiently and reliably. Yet, achieving natural, low-energy movement remains a major hurdle. Most existing methods either prioritize stability at the cost of energy efficiency or simplify robot dynamics so much that real-world performance suffers.
The new research, published in the Journal of Xi’an University of Science and Technology, proposes a solution that strikes a balance between these competing demands. By combining a sophisticated five-centroid inverted pendulum model with an innovative optimization strategy, the team has developed a gait-planning algorithm that enables robots to walk with greater efficiency than previously possible.
The core of the innovation lies in how the researchers model and optimize the robot’s body motion. Traditional approaches often assume a fixed torso height or allow only vertical movement, limiting the robot’s ability to adapt its gait dynamically. In contrast, Lu and his colleagues allow the robot’s body to move freely in three dimensions, enabling more natural and energy-saving motions. This flexibility, however, introduces a vast parameter space that is difficult to navigate using conventional optimization techniques.
To tackle this complexity, the team introduced a two-stage algorithm: gait parameter optimization (GPO) and gait synthesis (GSYN). The GPO phase operates offline, systematically exploring possible body trajectories using a discretized Fourier series representation. By dividing the motion space into a high-dimensional grid, the algorithm identifies a set of initial “seed” trajectories that meet basic stability criteria—specifically, ensuring the zero moment point (ZMP) remains within an allowable region on the foot.
The ZMP is a key concept in bipedal robotics. It represents the point on the ground where the net moment of all forces acting on the robot is zero. If the ZMP moves outside the footprint of the supporting foot, the robot risks tipping over. Most control strategies aim to keep the ZMP near the center of the foot for maximum stability. However, this conservative approach often leads to stiff, inefficient movement.
Lu’s team takes a more nuanced approach. Instead of forcing the ZMP to stay in the center, they define an allowable ZMP region (AZR) that permits the robot to shift its balance closer to the edge of the foot when appropriate. This allows for more dynamic, human-like motion and reduces the need for corrective torque, which in turn lowers energy consumption.
Once the seed trajectories are identified, the algorithm applies a gradient-based iterative method to refine them. Starting from each seed, it explores the local neighborhood in the parameter space, adjusting the body motion to minimize a carefully designed energy-consumption index. This index is based on the product of motor load torque and angular velocity—key factors that directly influence power draw in electric actuators.
What makes this optimization effective is its ability to converge reliably even in a high-dimensional space. The researchers found that after multiple iterations, the energy cost of walking dropped significantly—by as much as 18.66% on average compared to initial estimates. More importantly, the final gaits not only consumed less energy but also exhibited smoother ZMP trajectories, staying closer to the edge of the AZR without compromising stability.
After the optimal gaits are computed and stored in a database, the second phase—gait synthesis—takes over during real-time operation. When a robot is given a walking task, GSYN constructs a complete trajectory composed of starting steps, cyclic middle steps, and stopping steps. It retrieves pre-optimized joint-angle sequences from the database based on the desired step length and robustness requirements.
But the system doesn’t just replay stored motions. It incorporates real-time feedback from force-sensing resistors (FSRs) embedded in the robot’s feet. These sensors allow the robot to compute the actual ZMP during each step. If the measured ZMP deviates from expectations, the algorithm adjusts the allowable region for the next step using a PI (proportional-integral) control scheme. This closed-loop correction enhances robustness against modeling inaccuracies and environmental disturbances, such as uneven terrain or unexpected contact forces.
The result is a gait that is both energy-efficient and adaptable—capable of maintaining balance while minimizing unnecessary motor effort. In dynamic simulations, the algorithm outperformed two benchmark methods: one with a fixed body height and another allowing only vertical body motion. The energy savings were substantial, with reductions of up to 16.21% in real-world walking tests over a 90-centimeter distance.
One of the most compelling aspects of this work is its practical implications. While many gait-planning algorithms remain confined to simulation, Lu and his team validated their approach through both simulation and physical experiments. Their test robot, equipped with ten joints (five per leg), successfully executed the planned motions in real environments, demonstrating the algorithm’s readiness for real-world deployment.
The robot’s ability to adjust its gait on the fly based on sensory feedback is particularly promising for applications in unpredictable settings. For instance, in disaster zones such as collapsed mines—where the research was partly motivated—robots must navigate debris, slopes, and unstable surfaces. A rigid, pre-programmed gait would likely fail under such conditions. But with the adaptive, energy-aware control provided by this new algorithm, robots could operate longer on limited battery power while maintaining stability.
The choice of a five-centroid model also reflects a deeper understanding of robotic dynamics. Unlike simpler models that treat the entire body as a single point mass, this approach accounts for the distributed mass of the torso, upper legs, lower legs, and feet. This level of detail allows for more accurate torque calculations and, consequently, more realistic energy estimates. It also enables finer control over joint movements, reducing jerky motions that waste energy.
Another strength of the method is its computational efficiency in real-time operation. Although the offline optimization phase is computationally intensive—requiring extensive inverse dynamics calculations and iterative refinement—the online gait synthesis is lightweight. By pre-computing and storing optimal gaits, the robot can quickly retrieve and adapt them during operation, making the system suitable for embedded controllers with limited processing power.
The researchers also made deliberate design choices to enhance robustness. For example, they assumed the upper body remains upright during walking—a simplification supported by human biomechanics, as torso pitch angles during normal walking rarely exceed three degrees. They also assumed that the feet remain parallel to the ground, which is reasonable for most humanoid robots lacking articulated toes.
These assumptions reduce complexity without sacrificing performance, reflecting a pragmatic engineering mindset. Rather than chasing theoretical perfection, the team focused on creating a system that works reliably in practice. This balance between sophistication and practicality is a hallmark of effective robotics research.
The broader impact of this work extends beyond just walking efficiency. As robots become more integrated into society, their energy use will become a critical factor—both for operational longevity and environmental sustainability. A robot that consumes less power can operate longer between charges, reducing downtime and increasing utility. In remote or hazardous environments, where recharging may be difficult or impossible, energy efficiency can be the difference between mission success and failure.
Moreover, energy-efficient movement contributes to quieter, smoother operation—qualities that are essential for robots interacting with humans. Jerky, high-torque motions are not only wasteful but also intimidating. By minimizing unnecessary force and optimizing motion smoothness, this algorithm helps create robots that move in a more natural, human-friendly way.
The research also opens new possibilities for robot design. Traditionally, engineers have relied on powerful motors and heavy batteries to ensure sufficient performance. But with smarter control algorithms, it may be possible to use smaller, lighter components without sacrificing capability. This could lead to more agile, cost-effective robots that are easier to manufacture and deploy.
Looking ahead, the team suggests that their approach could be extended to more complex tasks, such as stair climbing, running, or carrying loads. The same principles of energy-aware optimization and real-time feedback could be applied to other forms of locomotion, potentially unlocking new levels of performance.
Furthermore, the modular nature of the algorithm—separating offline optimization from online synthesis—makes it adaptable to different robot platforms. With appropriate adjustments to the dynamic model and parameter space, the framework could be applied to robots of various sizes and configurations, from small educational platforms to large industrial machines.
The work also highlights the growing importance of interdisciplinary collaboration in robotics. The algorithm draws on concepts from mechanics, control theory, optimization, and biomechanics. It reflects a synthesis of ideas from multiple domains, demonstrating how progress in robotics often comes not from isolated breakthroughs but from the thoughtful integration of existing knowledge.
In an era where artificial intelligence and machine learning dominate headlines, this research serves as a reminder that classical control and modeling techniques still have immense value. While deep learning methods can generate impressive behaviors, they often require vast amounts of data and computational resources, and their decision-making processes can be opaque. In contrast, this physics-based approach offers transparency, predictability, and reliability—qualities that are essential for safety-critical applications.
That is not to say there are no limitations. The current implementation assumes flat, even terrain and does not yet handle dynamic obstacles or rapid changes in direction. Extending the algorithm to fully three-dimensional environments with variable footholds would be a logical next step. Additionally, while the energy model is more accurate than many alternatives, it still approximates real motor losses and does not account for all sources of inefficiency.
Nonetheless, the results represent a significant step forward. By rethinking how robots plan their steps—not just for stability, but for efficiency—the researchers have opened a new pathway for developing more capable, sustainable humanoid machines.
As robotics continues to evolve, innovations like this will play a crucial role in shaping the future of automation. Whether in search-and-rescue missions, elderly care, or industrial automation, robots that can move intelligently and efficiently will be better equipped to serve human needs.
The work by Zhiqiang Lu, Yuanbin Hou, Yun Meng, and Funa Zhou demonstrates that sometimes, the key to progress is not building more powerful machines, but teaching them to move with greater wisdom. Their algorithm doesn’t just make robots walk—it teaches them to walk smarter.
Energy-Efficient Gait Algorithm Helps Bipedal Robots Walk Smarter
Zhiqiang Lu, Yuanbin Hou, Yun Meng, Funa Zhou; Journal of Xi’an University of Science and Technology; DOI: 10.13800/j. cnki. xakjdxxb. 2021.0320