Shanghai Kepler Robotics Co., Ltd. announced a major update to the locomotion of its humanoid robot K2 “Bumblebee,” releasing what it describes as China’s first demonstration of a “hybrid-architecture disturbance-resistant” gait. In a video shared by the company, the robot maintains steady steps over bricks, plastic surfaces, and grass, and keeps its balance under external pushes. 

Kepler also unveiled a demonstration of its VLA+ training platform, which the company says allows the robot to interpret natural-language commands and execute a range of tasks. Together, the mobility and cognition updates are positioned for industrial and service deployments as Kepler pursues real-world use cases.

How the hybrid architecture works

K2 “Bumblebee” uses a hybrid actuation design that pairs planetary roller screw linear actuators with rotary actuators in a series-parallel arrangement. Kepler says this enables a human-like straight-knee gait while addressing precision and load requirements. In the scheme described by the company, the linear actuators act as the robot’s “leg muscles,” providing primary walking force, while the rotary units handle fine adjustments and gait switching to accommodate different terrains. 

The planetary roller screw converts rotary motion to linear motion through a planetary drive and threaded engagement; Kepler cites an energy conversion efficiency of 81.3%, along with low friction, high positioning accuracy, and higher load-bearing strength than ball screws. 

The company frames the hybrid approach as complex to model and control, but argues it yields more natural motion, improved energy efficiency, stronger task execution, and better adaptability.

Closing the sim-to-real gap

Kepler notes that achieving a straight-knee gait with this hybrid setup required solving both mechanical and control challenges, especially when moving from simulation to physical robots. According to the company, real-world factors such as dynamic inconsistencies, sensor noise and latency, actuator delays and nonlinearities, and variable ground properties can cause strategies optimized in simulation to slip or destabilize in practice. 

To address this, Kepler says it developed a hybrid series-parallel structure and conducted research on position, velocity, and torque dynamics between the roller screws and their paired joints. On the control side, the company reports combining reinforcement learning, imitation learning, and linear-actuator torque control so K2 can switch gaits dynamically across complex terrains and sustain a natural, human-like walk. 

Kepler also describes a “dual-data flywheel,” training the system with simulated data to boost general perception and language understanding, and with real-world datasets to learn directly from human movement. The layered VLA+ model is presented as supporting semantic recognition and task-oriented guidance for understanding, reasoning, and planning.

Toward commercial use

Kepler positions the latest gait upgrade as expanding K2’s applicability in commercial and industrial settings. Cited use cases include research and education, exhibitions and customer engagement, and data collection and training, as well as logistics and material handling, intelligent manufacturing, and specialized operations. 

The company highlights specifications aimed at enterprise environments, including one-hour charging for up to eight hours of operation, a 30-kilogram (66-pound) payload capacity, and a supply chain in which more than 80% of the hardware is developed in-house. 

Kepler also emphasizes cost efficiency and frames 2025 as a key year for humanoid robot commercialization, one in which ecosystems and scaled deployments, along with hardware and algorithms, will shape the market. The company portrays K2 “Bumblebee” as the first commercially available humanoid robot to integrate a hybrid architecture using roller screw linear actuators and argues this foundation positions it to compete on performance and deployment breadth. 

Kepler further situates its progress within China’s broader push in robotics, stating that domestic innovators have emerged as significant contributors to industry growth and that K2 aims to translate technology upgrades into practical, multi-scenario operations.