Unitree G1 Humanoid Catches Boxes of Up to 2 kg Thrown From 7 Meters Using Reflex

The Unitree G1 humanoid robot has learned to catch boxes weighing up to 2 kg thrown at it from up to 7 meters away. The result comes from a joint team at the University of Washington, USC, the National University of Singapore, UNC Chapel Hill and the Allen Institute for AI. The method is called Reflex. The robot sees the box only through an RGB-D camera on its torso, and it has about one second to do everything, which is how long the object spends in the air.

Why catching a box is harder than it looks

Most humanoid demos are built around predictable actions: walking, picking objects up from a table, dancing. A static object can be examined for as long as needed, but a flying one will not wait. Within a second, the system has to spot the box, estimate its position and velocity, predict its trajectory, and bring its hands and body to the interception point. An error of a few centimeters means the box hits the forearm and falls.

The setup has a notable constraint: the robot has no external motion-capture system and no sensors on the box itself. The only source of data about the object is the RGB-D camera mounted on the torso. It provides color and depth, but with noise and a limited frame rate. In a lab it is easy to lift this limitation with "external eyes," so the bet on onboard vision suggests the result is closer to real-world conditions.

What we know about the system

The published description gives the following parameters:

• platform: Unitree G1 humanoid robot;
• method: Reflex;
• box weight: up to 2 kg;
• throwing distance: up to 7 m;
• flight time: about 1 second;
• sensor: RGB-D camera on the torso, without an external mocap system.

The original announcement does not give the catch success rate, the control frequency, the policy architecture or the training approach (simulation, sim-to-real transfer, demonstration data). Without these figures, it is impossible to say how reliable the method is across different trajectories, lighting conditions and object shapes. These are worth checking in the paper and accompanying materials. Of particular interest are the success rate on long throws and the behavior when the camera is partially occluded.

Why the author list matters

The participants include universities with strong robotics and computer vision programs, plus the Allen Institute for AI, which is more often associated with language models and open research. That mix points to a trend: physical AI is ceasing to be the preserve of a few robotics labs. The platform is telling too. The Unitree G1 is a comparatively affordable humanoid, so other groups can reproduce the result on it more easily than on one-off robots costing hundreds of thousands of dollars.

Practical takeaway

For engineers and researchers, Reflex is interesting as a benchmark task: dynamic interception from a single camera within a second. It makes a handy test for your own perception and control stacks: if your system cannot handle a 2 kg box, it is a long way from working in a warehouse with falling packages.

For business, the picture calls for more caution. Catching objects in mid-air is so far a demonstration of fast reaction, not a ready-made logistics solution. Before any pilots, it is worth asking the authors three things: the success rate over a series of throws, the sensitivity to payload shape and mass, and the load on the robot's onboard compute. If the answers turn out to be convincing, the next step is obvious: combine such a reflex with task-level planning, where a model understands what to catch and why.