Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video

Dyna Robotics has launched Dyna-2, a world-action mannequin for robotic manipulation. It was pre-trained on a couple of million hours of selfish human video. That is roughly 170 years of steady waking expertise. Robot studying has been bottlenecked by action-labelled information, which teleoperation should intentionally produce. Dyna-2 checks whether or not odd human video can substitute. The analysis crew skilled a knowledge ladder from 1,000 to 1,000,000 hours and measured what scales. Three outcomes observe: a scaling legislation on human information, the primary switch of that legislation to unseen robotic information, and proof that video prediction drives the switch.

Is it deployable?

Yes, however as a vendor-operated system, not as downloadable weights. Dyna Robotics has introduced no public checkpoint, API, or license for Dyna-2. Deployment at present means shopping for a Dyna robotic cell, not self-hosting a mannequin.

• Which corporations: Dyna-1 robots already run in manufacturing in motels, eating places, and laundromats, per the corporate’s August 10, 2026 announcement. That factors at mid-market service operators and multi-site enterprises with repetitive, stationary manipulation work. It will not be a match for solo builders or analysis labs wanting native inference.
• Industries: Hospitality, business laundry, meals service, mild meeting and kitting, and amenities cleansing.
• Applications: The 14 post-training duties map cleanly to actual work: trash tray clearing, first-aid kitting, tote development, meals scooping, rope tying, hanger preparation, and focused drink retrieval from a fridge.

What is Dyna-2

Dyna-2 is a world-action mannequin (WAM): one generative mannequin that denoises future video and a future motion chunk, collectively or individually, on a video-diffusion spine. It was pre-trained on a couple of million hours of selfish human video, roughly 170 years of steady waking expertise.

Architecturally it’s a combination of transformers. Video and motion are tokenized individually and get distinct DiT layer stacks that attend to one another. Proprioception feeds straight into the motion transformer. Video tokens use causal masking; motion tokens use bidirectional self-attention and attend to context video tokens. Video tokens cross-attend to textual content, however textual content doesn’t straight affect motion tokens.

Training makes use of move matching. A video loss and an motion loss share a trunk as two separate marginal velocity fields. Because the motion community by no means takes the noised video latent as an argument, the coverage stays reactive at inference — it neither generates nor attends to predicted future video. The motion transformer is intentionally shallower and joins the video stream early, which the crew says improves real-time latency with out costing efficiency.

The three scaling outcomes

Dyna Robotics minimize nested subsets of precisely 1,000, 10,000, 100,000, and 1,000,000 hours, conserving an identical proportions from every supply. A bigger finances solely provides information, so curve variations can’t be attributed to distribution shift. A mounted, disjoint 100-hour validation set scores each rung.

1. A scaling legislation holds on human information to 1 million hours: All 4 metrics enhance monotonically and match energy legal guidelines: held-out MSE = 0.0691·D^-0.0184 (R²=0.919), [email protected] = 0.357·D^+0.0203 (R²=0.865). Across the ladder, [email protected] rises 51% in opposition to 12% for MSE.
2. That legislation transfers to robotic information the mannequin by no means noticed: The identical checkpoints have been scored zero-shot on 39 duties throughout two stationary bimanual YAM platforms — 12 inner, 27 from xdof ABC. Zero-shot motion MSE = 0.306·D^-0.0713 (R²=0.884). The crew stories an inflection between 10k and 100k hours.
3. The goal issues, and video is a separate axis: Joint denoising beat action-only on 39 of 39 duties at each motion scale. Holding action-labelled information mounted at 50,000 hours and including video-only hours drops zero-shot robotic MSE from 0.340 to 0.120. Notably, held-out human error doesn’t enhance — the profit of video is particularly cross-embodiment generalization.

On-robot outcomes

Each rung was post-trained on 14 duties, at most 10 hours of robotic information every, throughout three embodiments: 6-DOF YAM arms with parallel-jaw grippers, the identical arms with WUJI-2 20-DOF dexterous fingers, and a semi-humanoid prototype. Post-training used robotic information solely — no human-robot alignment, no co-training.

Mean normalized rating rose 20% → 28% → 45% → 53% throughout the ladder, finest on 9 of 14 duties at a million hours. Lockbox Key Turning is the edge case: 0% as much as 100,000 hours, then 90%. Bottle Cap Untwisting was post-trained on roughly 10 minutes of demonstrations and nonetheless climbed to 50%.

Against Dyna-1 — the corporate’s manufacturing VLA initialized from Qwen3-VL-4B — an early Dyna-2 reached 1.55× success fee and 1.12× grade, pooled over 7 duties and three checkpoints. At unseen buyer websites, Dyna-2 handed manufacturing standards 87% versus Dyna-1’s 46%, although each move close to 100% in home. A distillation pipeline additionally cuts video sampling from 10,203 ms to 110 ms on one H100.

Interactive explainer