In this conversation, we spoke with Fangzheng Wang, CEO of HopTo, about the constraints shaping hybrid robotic mobility, why hopping changes the energy model of small autonomous systems, and what would need to happen for hybrid locomotion to become a durable robotics category.
What made you decide that for small robots, flying and ground movement should be designed as one mobility system rather than treated as separate capabilities?
Versatility has always been an important criterion in the evaluation of robotic systems. In our case, the proposed robot can be regarded as a modest hardware extension of a conventional quadrotor platform. By simply mounting a passive compliant leg mechanism beneath a standard quadrotor and modifying the control algorithm accordingly, a conventional aerial vehicle can be transformed into a multifunctional robot capable of both flight and hopping. In the hopping mode, the robot can achieve enhanced endurance and payload capacity.
In your architecture, what is the main reason hopping can deliver better mission efficiency than staying airborne longer on a platform of this size?
For hopping robots, although they spend most of their time in the air, they do not need to generate continuous thrust to counteract gravity. Fundamentally, they are still terrestrial robots. In our robot, the rotors only need to compensate for the energy lost during the stance phase in the ascending phase of each hop to maintain a constant hopping height. In practice, the robot’s energy consumption can be reduced to about 25% of that of flight, resulting in nearly a fourfold improvement in endurance, as well as a significantly higher payload capacity.
How do you decide which problems should be solved by passive mechanics and which still need active control?
When we attempted to model the dynamics of the hopping robot, we found that the stance-phase dynamics are essentially a purely passive process. This is because the stance phase is extremely short, while the spring force can reach tens of times the robot’s body weight. Based on this observation, we deliberately treated the entire stance phase as a passive process and completely abandoned active control during ground contact.
Building on this idea, we developed a hopping controller based on the touchdown angle and achieved respectable control performance. Moreover, the passive stance-phase dynamics model also facilitates state inference from the robot’s behavior during stance, which in turn underpins our state estimation algorithm.
When turning Hopcopter from a research result into a repeatable system, which part was hardest to stabilize in practice: jumping behavior, landing recovery, leg reliability, or state estimation?
At present, the computational complexity of the state estimation algorithm remains relatively high, mainly because it requires continuous online optimization. As a result, we are currently unable to deploy the full state estimation pipeline on an MCU. Instead, a relatively high-performance SoC is needed to handle both state estimation and control, which inevitably increases the overall product cost. In future work, we plan to further optimize the algorithm and improve the system architecture to enable truly low-cost hopping robots.
What does reaction-force-based state estimation let you recover more reliably than a conventional IMU-first approach, especially during contact with uneven ground?
In fact, this is not an entirely new method developed independently of existing state estimation approaches. Rather, it is an improvement and customization built upon existing theories and methods, specifically tailored to the unique dynamics of hopping motion.
For hopping robots, conventional attitude estimation methods based on gravity sensing are no longer applicable, because the robot spends more than 95% of its time in free fall, during which acceleration measurements cannot be used to infer body orientation. Our method abandons acceleration-based attitude estimation and instead reconstructs the robot’s body orientation by observing its behavior during the stance phases across multiple hops. This enables drift-free attitude estimation for hopping locomotion.
In addition, the method also provides reliable velocity estimation, allowing the robot to achieve robust hopping control and stabilization using only proprioceptive sensors.
At this scale, weight, power, and robustness are all in tension. How do you make tradeoffs between sensing, onboard compute, structural protection, and mobility without weakening the core system?
The core challenge of Hopcopter lies in its control algorithm. As for sensing, onboard compute, structural protection, and mobility, the key is balancing the system around the control algorithm rather than optimizing any single component in isolation.
In real environments, what kind of terrain or obstacle pattern most clearly shows the advantage of hybrid locomotion over a standard micro quadrotor?
When robots are required to operate frequently in near-ground environments and repeatedly traverse tall obstacles, our hopping robot offers significant advantages in energy efficiency, endurance, and terrain traversability. It can readily perform tasks such as flying from the first floor to the second floor or clearing tall obstacles through high jumps.
These advantages mainly arise from its hybrid locomotion strategy and the flexible switching between its two modes. Compared with flying robots, our system provides better endurance because it does not rely on continuous thrust generation. Compared with ground robots, it offers superior obstacle-crossing capability and terrain adaptability.
In field conditions rather than lab conditions, what usually becomes the limiting factor first: control accuracy, mechanical tolerance, battery budget, or perception quality?
In real-world deployment, terrain uncertainty remains a critical challenge. Although our robot can adapt to a wide range of terrains, including flat and inclined surfaces, certain environments remain particularly difficult. For a single-legged robot, slipping can lead to severe performance degradation or even failure, while soft substrates such as loose sand may cause the robot to sink and prevent it from generating sufficient rebound.
Our current work primarily focuses on dynamics and control, and we have already achieved stable hopping on a subset of sufficiently rigid surfaces. In practical applications, however, the robot should also be capable of autonomously identifying regions unsuitable for hopping and actively avoiding them. Such a capability would significantly enhance the robot’s reliability in real-world scenarios.
At the same time, we recognize that falls are sometimes unavoidable for this type of robot. Therefore, the ability to recover rapidly after a fall is also essential. One of our ongoing research directions is to equip the robot with fast post-fall recovery capability. We believe this provides an alternative approach to mitigating the limitations imposed by terrain conditions on the robot’s locomotion performance.
When you evaluate a new use case, what tells you that HopTo is solving a real operational problem rather than demonstrating an impressive technical capability?
We spent a great deal of time thinking about how to transition from a laboratory prototype to a real product, and conducted extensive market research along the way. Our primary focus will be on the consumer market. Our five-year vision is to launch an AI-powered outdoor companion—leveraging Hopcopter’s exceptional mobility and extensibility, it can accompany people to a wide range of environments and offer a wealth of imaginative functionalities, bringing robotic companions from science fiction into reality. However, this will not happen overnight. Our product roadmap progresses from a remote-controlled toy, to a development platform, and ultimately to a fully intelligent robot. At this ICRA, we will be unveiling our first remote-controlled Hopcopter product, Hop-1.
You have worked in both sensing systems and robotic mobility. How has that shaped the way you think about the robot’s job in the field?
Hopcopter technology is well suited for three-dimensional spatial modeling, thanks to its capabilities in vertical movement, sustained continuous sampling, and visual payload capacity. Because it can thoroughly perceive and understand its surrounding environment, we believe that once equipped with an AI system, Hopcopter will be capable of highly effective intelligent autonomous control.
Many small robotics platforms are becoming more software-heavy. In your view, where does physical design still matter more than model sophistication in determining real-world performance?
Unlike traditional quadrotor robots with fully autonomous flight capability, our robot is required to interact with the environment frequently, which gives rise to fundamentally different requirements for environmental perception. At present, our state estimation and control framework relies entirely on proprioceptive sensors. This constitutes an elegant model-based approach to state estimation and control. These methods provide a solid foundation for the future development of more intelligent robotic systems. We believe that the ability of a robot to generalize and adapt to complex environments is crucial to the success of any robotic platform, and that this capability depends largely on software performance. Our current research is aimed at establishing a strong platform foundation for such future generalization capability.
Looking ahead, what kind of real-world adoption would make you feel HopTo helped define a lasting robotics category rather than just an inventive platform?
At present, we believe that this type of robot is particularly well suited for indoor industrial inspection tasks. Current indoor inspection operations generally rely on wheeled or quadrupedal robots. In contrast, our robot offers stronger adaptability to complex environments, enabling it to reach elevated locations that are inaccessible to conventional ground robots while maintaining a relatively compact form factor. In some cases, the robot does not need to use stairs to access other floors; instead, it can jump down from higher platforms or fly directly upward to elevated areas. This introduces a new solution for autonomous industrial inspection. We believe that, if deployed in such scenarios, our robot would have strong potential to achieve true commercial viability.
