UCSD's SteadyTray Robot Can Carry a Tray Without Dropping It

Humanoid robots are getting better at walking. Keeping stuff on a tray while they do it? That's harder.

Researchers at UCSD have built a system called ReST-RL that tackles exactly that problem. The approach decouples two things that usually get lumped together: how the robot moves, and how it keeps whatever it's carrying steady.

Most end-to-end learning systems try to do both at once. According to the researchers, that doesn't work well — the robot's natural gait creates oscillations that spill drinks. ReST-RL adds a separate "residual module" on top of a base locomotion policy. This module actively cancels out the wobble at the end-effector, essentially providing real-time counter-movements to keep the tray level.

In simulation, the system achieved a 96.9% success rate on variable velocity tracking and 74.5% robustness against external force disturbances. The researchers then deployed it on an actual Unitree G1 humanoid and reported reliable zero-shot sim-to-real transfer — meaning it worked on hardware without additional training.

That's the headline number. But here's the catch: these results come from a research benchmark called SteadyTray. The paper shows the robot can carry objects and handle some shoving. What it doesn't show is how this performs in a real hospital, restaurant, or warehouse over hours of continuous use. The gap between a controlled demo and a deployed system is where most robotics papers lose the thread.

The modular approach is sensible — separating locomotion from manipulation makes debugging easier. Whether it scales to uneven terrain, slippery surfaces, or a room full of people is still an open question.

The paper is on arXiv (2603.10306).

---

This article synthesizes the UCSD arXiv preprint on ReST-RL and SteadyTray, with verification of quantitative claims against the original source. The analysis notes the sim-to-real gap as the key uncertainty for real-world deployment.