Reflex Lets Unitree G1 Catch Tossed Boxes in About a Second

A research team has released the Reflex framework, which enables the Unitree G1 humanoid robot to perceive, predict the trajectory of, and coordinate its whole body to catch a box tossed by a human in about 1 second, using only an onboard RGB-D camera. Author @DJiafei relayed the team's view that for robots, "fighting looks easy, catching things is the hard part," since catching requires whole-body control, real-time reaction to object dynamics, and first-person perception.

Confirmed

• Real-robot tests covered small parcels, regular cartons, and large boxes, with successful cases across 0.5, 1, and 2 kg payloads; success rates varied with box size and load.
• Training followed three stages: 1) reinforcement learning for whole-body box-catching control; 2) training the policy to infer box dynamics from delayed and incomplete observations; 3) freezing the controller and training RGB-D observation histories to directly recover those dynamics. This decomposition gives each capability a direct training signal, making vision learning easier to scale.
• Ablations showed motion history is crucial: a single frame only reveals the box's position, while a sequence of observations is needed to infer where it's headed.
• Follow-up demos showed the full downstream pipeline: catch → carry → hand over → return, achieved by connecting Reflex to the pretrained SONIC walking controller.
• The project was led by Tao-Yang Jia (who is currently seeking PhD opportunities) with collaborators from multiple institutions, and both the code and paper are fully open source.

Why it matters

Catching dynamic objects tossed by humans is seen as a better showcase of integrated real-time perception and whole-body control than performative fighting. Reflex offers a scalable staged-training paradigm separating vision and control, fully open sourced for the community to reproduce and extend.

2026-10-09 ~ 2026-10-09 · 8 related posts

• [source] Reflex teaches a Unitree G1 humanoid to catch boxes thrown by humans — DJiafei · 2026-10-09
• How Reflex trains humanoids: 3-stage RL pipeline from control to RGB-D grounding — DJiafei · 2026-10-09
• Reflex real-world tests: catching parcels to large cartons up to 2 kg payloads — DJiafei · 2026-10-09
• Reflex simulation: 85.7% catch success from RGB-D vs 87.9% privileged — DJiafei · 2026-10-09
• Reflex hits 85.7% catch success from RGB-D and chains catch-carry-handoff demos — DJiafei · 2026-10-09
• [source] Reflex demo: catch, carry, hand over — code and paper open-sourced — DJiafei · 2026-10-09
• [source] UW CSE robotics unveils its first humanoid project, built for real factory work — DJiafei · 2026-10-09

1 near-duplicate retellings: DJiafei