← Leaderboard

MachEmbodied-U0

World-action
Li Auto · Sep 22, 2026

Li Auto’s unified world-action model that jointly predicts video and actions with flow matching, pretrained on about 4,200 hours of robot and egocentric data. It reports LIBERO, LIBERO-Plus and RoboDojo; its real-world results are qualitative only.

Type
World-action
Parameters
—
Backbone
Lance (Mixture-of-Transformers understanding + generation experts)
Weights
Open
License
Apache-2.0
Commercial use
Allowed
Released
Sep 22, 2026
Robots
Franka Panda (sim), ARX X5 (sim)
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

3 of 13 benchmarks
#29 of 5411.5%
LIBERO-PlusSimulation
#13 of 2883.1%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#9 of 7099.0%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks
Robustness drop: 99.0% on LIBERO → 83.1% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.
Wrong or missing number? Report an error