← Leaderboard

ScaleVLA

VLA
Galbot et al. · Nov 16, 2026

Technical report from Tsinghua, PKU, SJTU and Galbot on scaling a VLA to a 552B mixture-of-experts LLM backbone with an action DiT trained from scratch. It reports LIBERO plus partial LIBERO-PRO and RoboCasa results; its real-robot result is 9/10 in-domain on a RoboDojo-style plug-in-charger task.

Type
VLA
Parameters
552B
Backbone
Frozen DeepSeek-V4.1-Flash (MoE) + GR00T N1.7-style action DiT
Weights
Unknown
License
–
Commercial use
Unclear
Released
Nov 16, 2026
Robots
Franka Panda (sim)
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

1 of 13 benchmarks
LIBEROSaturatedSimulation
#47 of 7096.1%, self-reported by the authors, not a leaderboard entry
Not reported on 12 benchmarks
Wrong or missing number? Report an error