2 Frames of Delay Drops VLA Ball-Catching Success from 64% to 0%
wzenus · x · 2026-10-22
The team benchmarked OpenVLA, π0.5, and GR00T under tiny real-time delays with striking results:
- On Flappy Bird, a 1-frame delay leaves just 4.2% of the zero-delay reward
- In fast ball-catching, a 2-frame delay drops success from 64% to 0%
The explanation: agents act on stale states, and latency is hard to reproduce in training—random network spikes can't be batched away, and large-batch training slows individual episodes for throughput. The takeaway: faster inference alone doesn't solve real-time control; agents need to be latency-aware by design and training.
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