1X World Model (1XWM) 정책

1X Technologies · 2026-01

기타New in 2026Video backbone 14Bvideo world model

A video-generation-based policy in which an IDM extracts actions between video frames generated by a world model.

Paper

Backbone LLM: None (video diffusion backbone)
Vision encoder: —
Action head: Inverse dynamics model (IDM) (video world model)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): 14B text-conditioned generative video diffusion backbone + inverse dynamics model (IDM: Depth Anything backbone + flow matching head, 8-frame sliding window)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed (embodiment fine-tuning on 70 hours of NEO robot data)
Data type: Web vision-language (text-conditioned web video) + human video (egocentric) + real robot (NEO logs)
Data scale: 900 hours of egocentric human video; 70 hours of NEO robot; 400 hours of unfiltered robot data for IDM; web video scale not disclosed
Main datasets: In-house egocentric video, NEO sensorimotor logs
Tasks & evaluation: 30 trials per task: in-distribution chip picking about 100%·watering can about 80%, out-of-distribution cereal pouring 0%·drawing 0%
Robot: 1X NEO

Inference device & speed: 5-second video generation about 11 s + IDM about 1 s, multi-GPU inference
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Multi-GPU server

Reinforcement learning: Not used · Web video pretraining → human egocentric video mid-training → NEO log fine-tuning (supervised). Learning from its own experience is a future goal
RL evidence: https://www.1x.tech/discover/world-model-self-learning
Base control: Fixed · No mention of locomotion·movement (IDM extracts action trajectories from generated video, bimanual manipulation tasks)

Key features: A video-generation-based policy in which an IDM extracts actions between video frames generated by a world model. Best-of-8 parallel generation raised tissue-pulling success rate from 30% to 45%.
Notes: 1X 2026-01-12, TechCrunch 2026-01-13 (secondary). Data hours (h) is the largest single item (900 hours of egocentric video). Whether the backbone is a Transformer is not disclosed (note classifies it for comparison). 98.5% of the post-training mix is pick-and-place.
Verification: Confirmed