← Leaderboard

Hume

VLA
SJTU & Shanghai AI Lab · May 27, 2025 · built on π0

Dual-system VLA from SJTU, Shanghai AI Lab and AgiBot that samples candidate action chunks, scores them with a value head and refines them with a fast denoising system. Its pretrained weights were exported from π0.

Type
VLA
Parameters
—
Backbone
π0 + value-guided System-2 thinking
Weights
Open
License
MIT
Commercial use
Allowed
Released
May 27, 2025
Robots
WidowX, AgiBot G-1, Franka, Franka Panda (sim)
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

2 of 13 benchmarks
LIBEROSaturatedSimulation
#17 of 7098.6%, self-reported by the authors, not a leaderboard entry
SimplerEnv WidowXSaturatedSimulation
#17 of 3060.8%, self-reported by the authors, not a leaderboard entry
Not reported on 11 benchmarks
Wrong or missing number? Report an error