Skip to content

Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models

Session: Presented in Dyna Robotics' talk on Day 4. See the Day 4 notes. Report: Dyna Robotics, August 2026: dyna.co/dyna-2. The follow-up, Dyna-2.1 (Sep 29, 2026), is at dyna.co/dyna-2.1.

Summary (from Dyna's public report pages, read through a web summary; the full report is not yet read closely)

  • Claim: A world-action model pre-trained on more than one million hours of egocentric human video. It shows a scaling law on held-out human data and, Dyna says, the first human-to-robot transfer scaling law.
  • Data: Egocentric human video. Clips that pass a quality bar carry 3D hand-pose tracks; the rest have no action labels. The scaling-law variants are pre-trained on human data only. Dyna notes its production models use different recipes, and Dyna-2.1 mixes in robot data from its fleet.
  • Human-to-robot transfer, zero-shot on 39 robot tasks never seen in pre-training:
Pre-training human hours 1k 10k 100k 1M
Action MSE ↓ 0.180 0.174 0.124 0.117
Accuracy@0.5 ↑ 0.067 0.074 0.136 0.159
  • Fits: MSE = 0.306 · D−0.0713 (R² = 0.884), and acc = 0.0241 · D0.139 (R² = 0.918), where D is human hours.
  • Post-training: "a few hours of robot data" per task. One example: 10 minutes of teleoperation data to fine-tune bottle-cap opening.
  • In deployment: 87% of customer acceptance tests passed zero-shot, against 46% for Dyna-1. Napkin folding at 95 an hour with 93% quality.
  • In Dyna's stack: It runs as the policy layer at about 5 Hz, between a VLM workflow orchestrator and an RL whole-body controller at 100 Hz, on the Taku robot.

Questions / discussion

  • How does this compare with DreamZero, which transfers from 12 minutes of human video, and Being-H0.5, which pre-trains on 16K human hours plus robot data?
  • The curve is flat from 1k to 10k hours, then improves sharply by 100k. Is there a data threshold?