A team led by Yeonseo Lee introduced TempoBridge, a lightweight add‑on for vision‑language‑action (VLA) robot models. It reads tempo cues like “quickly” or “slowly” from the instruction and adjusts motion speed on the fly. The method uses frozen VLA representations, so it needs no extra tempo‑specific demonstrations or fine‑tuning. In tests on the LIBERO benchmark, Tempo Success Rate rose from 52.6% to 89.7% while overall task success stayed high. Real‑world robot trials also showed the system could follow tempo commands.
Why it matters
It lets end users control how quickly a robot works just by saying “fast” or “slow”, without needing new training data.
