KAIST's VSpark Speeds Up Autoregressive VLA Robot Models 2.1x with Speculative Decoding
kaist-ai · hf · 2026-11-20
KAIST researchers present VSpark, bringing speculative decoding to autoregressive vision-language-action (VLA) models with lossless 1.28x–2.10x speedups on robot manipulation benchmarks.
- Three key ideas: calibrated action-noise rollouts, hidden-feature alignment between draft and target, and reuse of rejected-draft continuations—since different action token sequences can encode similar robot motions, predictions from rejected branches often remain valid.
- On LIBERO, VSpark reaches 1.93x on π₀-FAST and 2.10x on ActionCodec while retaining exact target verification.
A practical step toward lower-latency robot policy inference.
More from Embodied
- AI writes all code to make a real robotic arm play chess, open-sourced — FinanceYF5 · 2026-11-20
- WAM-Policy Brings Language-Conditioned Loco-Manipulation to Arm-Free Quadruped Robots — Maincode · 2026-11-20
- RoboColosseum Aims to Benchmark Robot Policies in the Real World at Unprecedented Scale — DJiafei · 2026-11-20
- Dev Tours Seeed Studio's Shenzhen Factory: PCBA Lines and Robot Arm Assembly Up Close — haydendevs · 2026-11-20
- Musk says Optimus robotic hands could be given to people in need — CyberRobooo · 2026-11-20
- Zeromatter opens its pre-name pitch deck 3.5 years after pitching simulation infra for physical intelligence — soleio · 2026-11-20