Researchers led by Tianqi Gong introduced GTI, a method that keeps a robot’s overall task intent grounded while it works through many steps. GTI looks at what the robot sees and what it has already done to decide which subtask is active and what visual state should be reached next. In real‑world robot trials, GTI lifted average full‑task success from 21.7% to 59.2%, more than doubling performance.
The paper also reports gains on simulated benchmarks, with success on RoboCasa365 tasks rising to 31.8% and overall success on VLA‑Arena reaching 47.6%.
Why it matters
Higher success rates mean robots can complete multi‑step chores more reliably for industry users.