DreamZero

NVIDIA GEAR · 2026-02

기타New in 202614Bvideo world model

A world action model (WAM) that jointly predicts video and actions.

Paper

Backbone LLM: None (video diffusion backbone)
Vision encoder: —
Action head: Joint video + action flow matching prediction (video world model)
Chunk: 48 steps (1.6 s)
Control rate: 30 Hz
Architecture (original summary): Wan2.1-I2V-14B-480P video diffusion → autoregressive DiT (flow matching·teacher forcing); relative joint position actions, 30 Hz, 48-step chunk (1.6 s)

Pretraining accelerators: Not disclosed (no training-GPU statement; H100·GB200 mentioned only for inference)
Pretraining duration: Not disclosed
Pretraining steps & batch: 100K steps each for AgiBot · DROID, global batch 128
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed; adapting to a new embodiment takes 30 min of play data, 10–20 min of other-robot·human-video-only data
Data type: Real-robot teleoperation + video-pretrained backbone (Wan2.1)
Data scale: About 500 hours of AgiBot G1 teleoperation (22 environments, 7.2K episodes) + DROID public dataset
Main datasets: AgiBot G1 teleoperation, DROID
Tasks & evaluation: AgiBot G1 trained-task progress 62.2% (best pretrained VLA 27.4%), unseen 39.5% (16.3%); DROID-Franka progress 49% · success 22.5%; 1st on MolmoSpaces · RoboArena
Robot: AgiBot G1, Franka(DROID)

Inference device & speed: 5.7 s→150 ms per chunk (38× on GB200 including DreamZero-Flash; system optimizations alone about 9× on H100, about 16× on GB200), about 7 Hz closed loop; README GB200 about 0.6 s, H100 about 3 s, 2 or more GPUs
Inference memory: Not disclosed
Weights bf16 (calc.): 28 GB
Where it runs: 2 or more GPUs (README)

Reinforcement learning: Not used · flow matching·teacher forcing supervised learning
Base control: Fixed · No base motion stated in evaluation tasks (AgiBot G1 relative joint position actions; training data includes navigation)

Key features: A world action model (WAM) that jointly predicts video and actions. With 39.5% progress on unseen tasks it beat pretrained VLAs (16.3%), and it is the foundational work for GR00T N2.
Notes: arXiv 2026-02-17. DreamZero-DROID 14B·DreamZero-AgiBot checkpoints (Apache 2.0). Data hours (h) for AgiBot only (DROID separate). Speedup conditions clarified per verification notes. NVIDIA introduced GR00T N2 as 'based on DreamZero research' (2026-03-16).
Verification: Confirmed