AGILE2.0 Integrated Sensing-Control Model Ushers Robots into 'Eyes-Open' Motion Era
Recently, Zhiyuan released AGILE2.0 (AGIBOT Generative Intelligent Locomotion Engine), an integrated sensing-control model that allows robots to see, think, and move simultaneously, achieving natural and flexible motion akin to humans. To demonstrate the robot's visual capability, Zhiyuan produced a creative short film titled "Acrobatic BOT," moving beyond conventional technical videos to vividly present this revolutionary motion intelligence.
The film opens with a dramatic fire hoop jump: the Lingxi X2 robot locks onto the hoop's center and leaps through it effortlessly. Subsequent stunts include diabolo spinning, three-person jump rope, double leapfrog, collaborative platform climbing, juggling three balls, and cooperative box moving—all targeting the core challenge of humanoid robotics: enabling robots to use their eyes to see, then plan and act. The high-difficulty visual-driven dynamic coordination tasks validate the breakthrough capabilities of the AGILE2.0 visual end-to-end model.
The fundamental difficulty lies not in robot vision but in making robots act based on what they see. Humans naturally synchronize eyes and body: eyes capture changes real-time, the brain predicts delays, and limbs adjust smoothly. Robots, by contrast, are separate systems: cameras capture images and AI computes target positions, introducing time lags, while motor responses occur in milliseconds. If lagging visual signals directly guide motion, robots tend to shake or become unstable. To address this, AGILE1.0 first established a perception-motion link, but parallel mobile manipulation lacked whole-body coordination and dynamic coupling. Most industry solutions still simply combine "mobile + manipulation" without unified dynamics, leading to instability in extreme balance, high-speed interaction, and fine manipulation.
AGILE2.0 builds on 1.0 to achieve deep coupling across environment observation, terrain understanding, dynamic traversability analysis, whole-body motion control, contact-state switching, and end-effector fine manipulation. It forms a truly end-to-end, integrated mechanism without intermediate modules or manual stage division, realizing unified closed-loop coupling and autonomous reasoning for vision, environment comprehension, whole-body control, and coordinated upper-lower limb tasks. The highlight is the "ball walking" scene in "Acrobatic BOT," where the Lingxi X2 stands on a large ball and rolls it forward, displaying extraordinary dynamic balance—an industry first for a bipedal humanoid robot autonomously performing this feat, signaling that humanoid motion intelligence has reached human-comparable levels.
In the jump-roping scene, two Lingxi X2 robots swing a long rope rhythmically while a third watches, times its entry, and jumps in gracefully, then exits. In the box-moving scene, one robot picks up a box, and another sees it, knocks the box down (as if saying "Let me show you"), then redeems it by stacking the box perfectly atop another. These tasks not only demonstrate exceptional integrated sensing-control ability but also showcase whole-body loco-manipulation—performing fine operations during dynamic movement while maintaining stability and continuous motion.
The industrial value of AGILE2.0 lies in its visual end-to-end closed-loop sensing-control architecture. For extreme scenarios with rapid environmental changes and dynamic obstacles, AGILE2.0 enables real-time visual closed-loop reasoning and extreme dynamic constraint optimization. For two-robot or multi-robot collaboration, it establishes cross-robot state sensing and whole-body behavioral coordination. When humans randomly enter the workspace or unexpected interactions occur, the model ensures safe coexistence and autonomous recovery.
The release of AGILE2.0 breaks the limitation that robots can only work in structured environments. It effectively adapts to dynamic disturbances, human-robot coexistence, and multi-robot collaboration in real complex scenes, reducing the cost of customization for real-world deployment. This provides a stronger motion intelligence foundation for humanoid robots entering the physical world, accelerating the scalable, commercialized application of embodied AI in deployment scenarios.