Reflex lets a Unitree G1 catch boxes people throw at it
The headline 85.7% catch rate is in simulation. On hardware, the robot caught 8 of 20 throws aimed to the side.
Jiafei Duan and collaborators released Reflex, a whole-body framework that lets a Unitree G1 humanoid catch boxes thrown at it by a person, using a single onboard RGB-D camera (color plus depth). The robot has about one second to see the box, predict where it is going, and move its legs, torso and arms into place.
Humanoid fighting looks easy. Humanoid catching is hard.
A single frame shows where the box is; a sequence helps infer where it is going.
The pitch is logistics. Duan frames catching tossed boxes as the thing humans already do when unloading trucks or moving packages, and the thing humanoids should do faster and more precisely.
The numbers
In simulation, Reflex hits 85.7% catch success from onboard RGB-D, against 87.9% when the policy is handed privileged box observations it could not actually see. That gap is small, which is the interesting part, since it suggests the camera-only version is not giving up much.
On the real robot the only figure given is narrower. Reflex can take a step to intercept throws that are not aimed at its chest, and in hardware evaluation it caught 8 of 20 throws aimed to either side. The team also tested small parcels, regular boxes and large cartons with payloads of 0.5, 1 and 2 kg, and says success rates vary by box size and payload. The clips posted show successful examples across those conditions, not full success rates.
How it is trained
Three stages. First, learn whole-body catching with reinforcement learning. Second, infer box dynamics from delayed, incomplete observations. Third, learn that dynamics representation from a history of RGB-D frames while keeping the controller frozen.
That history matters more than you might expect. In simulation, removing visual history cuts catch success by 35.6 percentage points on centered throws and 39.8 points on throws offset by 0.6 m. One frame tells you where the box is. A sequence tells you where it is going.
The team also chained Reflex to the pretrained SONIC walking controller for a longer demo, catch, carry, hand over, return.
Who built it
Duan credits @TaoyangJia as project lead, with @weikaih04, @linxins2, Jared Darlington, @jieyuzhao11, @yuewang314, @RenZhongzheng and @RanjayKrishna. He calls it the first humanoid project from @uwcse robotics. Code, paper and interactive simulation replays are linked from the thread.
Shown versus claimed. The 85.7% is simulation. The hardware evidence in public right now is 8 of 20 on deliberately off-target throws plus a set of successful clips, so nobody outside the team has a real-world success rate across the easy cases yet.