Blog

kernel.queue([lift_left,lift_right])# while it thinks 3.3×faster667 s → 204 sFewer round trips164 → 59 requests1.8×Pipeliningmoves whileit thinks1× vanilla loop

Real-time RUA·Nov 2026

Getting Robot-Use Agents to Run in Real Time

A real two-arm robot finishes the same task 3.3× faster: fewer round trips, and motion while the model thinks.

r = tools.return_home(arms=[a])if not r.ok: 39.9%on RoboDojo5.2% in the October previewCode mode, planned motionOctoberv0.1RoboDojo39.9RoboTwin79.6RoboLab79.2LIBERO-Pro88.0

ORA·Nov 2026

v0.1: Robot-use agents now write code, move smoothly, and reach 40% on RoboDojo

Code mode, planned motion, and six open models in one harness.

vision_ground("ketchup bottle")└ T0 score 0.69└ T1 score 0.67 69.2%on RoboTwinzero-shot, open weightsPerception and motion toolsno VLA, no robot training69.2RPent + GPT-5.562.4π0.547.9

ORA·Oct 2026

Introducing : Open Robot-Use Agents Zero-Shot Solve Robot Tasks

An open harness of perception and motion tools: no VLA, no robot training.

You can also read our lab's blog.