LAGEN: Real GPU Latency Leaves 36 VLA Policies at Just 6.6% of Their Returns
wzenus · x · 2026-10-22
A Northwestern/Stanford/NVIDIA collaboration launched LAGEN, asking whether vision-language agents are latency aware:
- In real time on an RTX 3090, the median return of 36 VLA policies is only 6.6% of their no-latency return; adding just two frames of delay drops Flappy Bird and Demon Attack returns to 1.2–7.1%
- Standard evaluations that pause the environment hide latency effects; LAGEN ships profiling, replay, and training tools to measure deployment latency and train agents to cope with it
- They also probe general models: GPT "Sol" playing Asterix by watching the screen, no code, outscores the trained OpenVLA zero-shot under the same latency profile (5.8k vs 3.6k)
Paper, code, data, and hardware latency profiles are all publicly released.
More from Embodied
- Robotic bricklaying systems reportedly lay 1,000+ bricks per hour vs 300-500 a day for humans — 2C_ornot2C · 2026-10-22
- HKU-led team launches Embodied Playground, 410-task benchmark for model physical intelligence — YiMaTweets · 2026-10-22
- AgenticRoboRL lifts a 9B robot planner from 18% to 71% on LIBERO-PRO — UCF-Ai-LLMs · 2026-10-22
- Berkeley and NVIDIA's AW-0 couples code, video, and action streams for robot manipulation — Zhuoyang Liu · 2026-10-22
- Yao Ming takes the stage to talk robotics at a tech event — micoolcho · 2026-10-22
- Matt Wolfe demos 'personal bodyguard robot' standing in a junkyard — Matt Wolfe · 2026-10-22