DYNA-2

Dyna Robotics · 2026-08

기타New in 2026Not disclosedvideo world model

A world action model (WAM) pretrained on over 1M hours of egocentric human video.

Paper

Backbone LLM: None (video diffusion backbone)
Vision encoder: —
Action head: Separate video·action flow matching (video world model)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): MoT (DiT layers) on a video diffusion backbone: causally masked video tokens + bidirectional action tokens, text cross-attends only to video tokens; separate flow matching velocity fields for video·actions; pseudo-actions from 3D hand pose tracking (wrist pose·grip signal)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed; fine-tuned with about 10 min of teleoperation for the bottle-cap task (secondary: 13 min), on-site local fine-tuning a few hours (secondary)
Data type: Human video (egocentric) pretraining + real-robot teleoperation post-training
Data scale: Over 1M hours of egocentric human video (scaled with nested subsets of 1K · 10K · 100K · 1M hours); no robot frames in pretraining (secondary)
Main datasets: In-house egocentric human video
Tasks & evaluation: 14-task average normalized performance 1K 20% · 10K 28% · 100K 45% · 1M hours 53%; language following 96%; zero-shot pass at customer sites Dyna-1 46% · Dyna-2 87%; early Dyna-2 vs Dyna-1 head-to-head 65% vs 29% (6% ties); pooled WAM success rate is 1.55× that of VLA
Robot: Bimanual YAM (12 internal + 27 external xdof ABC tasks)

Inference device & speed: 1-step video generation distillation cuts 3-second video latent generation from 10,203 ms to 110 ms (H100, about 90×)
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Not disclosed

Reinforcement learning: Not disclosed · Training method details not disclosed
Base control: Fixed · No base control (stationary bimanual YAM; no mention of locomotion for the semi-humanoid prototype either)

Key features: A world action model (WAM) pretrained on over 1M hours of egocentric human video. Robot normalized performance rose from 20% to 53% as pretraining grew from 1K to 1M hours.
Notes: Corrections applied: the note's 'WAM 65% vs VLA 29%' is, per the source, an early Dyna-2 vs Dyna-1 head-to-head (6% ties). The pooled result in the same paragraph is WAM at 1.55× VLA's success rate·1.12× its grade. Official page 'August 2026' (secondary: August 10·11). All metrics self-reported. Scaling fit: human MSE 0.0691·D^-0.0184 (R² 0.919).
Verification: Corrections applied