← Leaderboard

Cosmos Policy

World-action
NVIDIA · Jan 22, 2026

Robot policy fine-tuned from the Cosmos-Predict2 video model, reported on LIBERO, the classic 24-task RoboCasa and real ALOHA tasks.

Type
World-action
Parameters
2B
Backbone
Cosmos-Predict2-2B
Weights
Open
License
NVIDIA One-Way Noncommercial License
Commercial use
Not allowed
Released
Jan 22, 2026
Robots
Franka Panda (sim), ALOHA
SO-101
–
LeRobot
–

Benchmark results

2 of 13 benchmarks
#12 of 2483.1%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#17 of 4698.5%, self-reported by the authors, not a leaderboard entry
Not reported on 11 benchmarks
Robustness drop: 98.5% on LIBERO → 83.1% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.

Adoption

updated Nov 17, 2026
GitHub stars
898
GitHub forks
109

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
H100–1 denoising step (RoboCasa)160 ms1–32–Jan 22, 2026 ↗
unspecified–LIBERO sim––6.8 GB––Jan 23, 2026 ↗

Trained on

Runs on

Wrong or missing number? Report an error