Traditional robots work like this: first perceive, then plan, then control. Three stages executing serially, each waiting for the previous one to finish. This pipeline has run for decades — stable, but slow. Especially in scenarios requiring rapid reactions, the latency between perception and control is enough for the robot to miss the optimal action window.

Zhiyuan Robotics' AGILE 2.0 model, released in September, proposes a different approach: perception is control; seeing is acting. It merges visual perception and motor control into a single end-to-end step. The robot doesn't need to "think it through before moving" — it "moves as it sees." Zhiyuan calls this "eye-open motion."

"Eye-Open Motion": From Serial Pipeline to Parallel Response

The concept of "eye-open motion" is intuitive. Imagine catching a ball thrown at you — you don't first analyze the trajectory, calculate the landing point, plan the arm movement path, and then reach out. You reach while watching. Perception and motion are nearly simultaneous. The human nervous system works this way: visual signals reach the brain while motor commands are already being dispatched.

AGILE 2.0 attempts to replicate this capability in robots. Its core is an end-to-end unified perception-control model — input is raw visual signals from cameras, output is directly joint motor commands, with no explicit perceive-plan-control steps in between. The model has learned to see and move at the same time.

In the LingXi X2 robot's live demonstration, the system showed real-time responsiveness in dynamic environments: reaction speed to obstacles during movement, precision in grasping target objects, and adaptability to unexpected disturbances — all significantly improved compared to traditional serial pipelines.

Why End-to-End Is a Key Turning Point for Embodied AI

End-to-end isn't a new concept — it's been extensively discussed and validated in autonomous driving. But in embodied AI, end-to-end deployment has faced a core challenge: insufficient training data.

Autonomous driving has massive road-driving data for training end-to-end models. But robot manipulation data is far from enough — every robot morphology is different, every task scenario is different, and real-world interaction data collection costs are extremely high. Over the past few years, robotics advances have mainly concentrated on training policies in simulation environments (like Isaac Sim) then transferring to the real world — but the sim-to-real gap remains a persistent problem.

AGILE 2.0's breakthrough lies in finding a more efficient way to leverage visual data. By fusing perception and control into a single model, it reduces intermediate information loss. In traditional pipelines, the perception module outputs abstract features, the planning module makes decisions based on these features, and information is lost at each conversion. End-to-end models preserve all information from raw visual signals, letting motor control directly "see" the details.

Is Embodied AI's "iPhone Moment" Still Far Away?

Zhiyuan isn't the only company working on end-to-end perception-control unification. Globally, Tesla's Optimus team, Figure, 1X Technologies, and domestically Unitree are all exploring similar directions. But each has a different entry point: some focus on hardware, some on simulation, some on foundation models.

AGILE 2.0 represents a model-side breakthrough — it doesn't change hardware or rely on specific simulation environments, but improves robots' real-time responsiveness through model architecture innovation. This approach's advantage is greater generality, deployable across different robot platforms.

Of course, "eye-open motion" has only been validated in specific scenarios so far. From demo to actual industrial deployment, there's still a mountain of reliability, safety, and generalization problems to solve. But the direction is right: when a robot's perception and control are no longer two independent modules but a unified whole, the ceiling for embodied AI jumps by an order of magnitude.

AGILE 2.0 may not be the ultimate answer, but it poses the right question: why make robots "think before they act"? Why not let them "move as they see"?

See you tomorrow.