AgiBot AGILE 2.0: How Its Robots Aim to Move and Manipulate at Once

AgiBot describes AGILE 2.0 as a sense–control locomotion model that connects what a robot sees with how it moves and makes contact. The goal is to keep those decisions linked as conditions change, so the robot can handle an object while moving rather than treating walking and manipulation as separate steps. The reported Lingxi X2 demonstrations show that idea in action, but leave its reliability outside staged settings unresolved. 513

What changed from AGILE 1.0?

AgiBot says AGILE 1.0 connected perception with movement. AGILE 2.0 is intended to couple more of the task in one end-to-end loop: observing the environment, interpreting terrain and changing traversability, controlling whole-body motion, switching contact states and directing precise end-effector actions. That contrasts with a staged “walk, stop, then manipulate” approach, in which locomotion and object handling are planned as distinct phases. These are descriptions of AgiBot’s proposed architecture, not evidence that every internal control step has been eliminated. 5611

How does the visual feedback loop work?

The key distinction is updating actions while moving. As new visual observations arrive, AGILE 2.0 is meant to revise terrain and movement decisions alongside the actions needed to reach, carry or maintain contact with an object. In AgiBot’s account, those decisions feed into coordinated whole-body control instead of requiring the robot to finish walking before it begins handling something. Foot-contact feedback also matters for balance in the rolling-ball demonstration; “visual” does not mean vision is the robot’s only input. 51018

That coupling could be useful when the ground, an object or another participant moves unexpectedly. It should not be confused with proven safe operation around people: reports describe adaptation and recovery as intended capabilities, but the cited demonstrations do not quantify how consistently they work. 519

What do the Lingxi X2 demonstrations establish?

In AgiBot’s Acrobat BOT presentation, Lingxi X2 robots are reported walking on a rolling ball, jumping a long rope and cooperatively carrying a box. The acts illustrate different control challenges: maintaining balance as contact changes, timing movement to a moving rope and coordinating carrying with locomotion and another robot. They show the kind of behavior AGILE 2.0 targets, not success rates across repeated trials or proof that the same skills transfer unchanged to workplaces. 101318

AgiBot has also used G2 robots in tablet-handling work at a factory; one account describes an eight-robot, six-day trial. That is evidence about a different AgiBot deployment, not an AGILE 2.0 field test on Lingxi X2. 29

Is AGILE 2.0 the same as GE-Act 2.0?

No. AGILE 2.0 is presented as a model for tightly coupled locomotion and whole-body interaction. GE-Act 2.0 is a separately described world–action model for robotic manipulation, built around a control-oriented visual encoder, a single-step visual planner and an inverse-dynamics model. Results for GE-Act’s manipulation tasks cannot, by themselves, validate AGILE 2.0’s moving-and-handling claims. 51

What evidence is still missing?

The decisive test would be repeated, independently observed work on unfamiliar terrain and tasks, with reported failures as well as successes. Comparisons under the same conditions against AGILE 1.0 and a staged walk-then-manipulate controller would clarify what the tighter coupling adds. Trials should also measure recovery from slips, obstructed views and unexpected motion around the robot. Until such results are available, the Lingxi X2 acts are best read as demonstrations of a promising control approach—not a measure of dependable deployment performance. 513