DYNA-2: The Robot Foundation Model Trained on 1 Million Hours of Human Video

Training Objectives

During pre-training, DYNA-2 used a video prediction objective — learning to predict future video frames from past observations — combined with a video co-training algorithm that jointly learns visual representations and action prediction . This approach contrasts with the VLA architecture of DYNA-1, which directly maps perception to robot actions without intermediate video prediction .

The company demonstrated scaling across four orders of magnitude, from 1,000 hours up to 1,000,000 hours of pre-training data .

Performance Improvements Over DYNA-1

The company ran direct head-to-head comparisons between DYNA-2 and DYNA-1 under matched training conditions and real-world deployments.

Head-to-Head Benchmark Testing (7 tasks)

• DYNA-2 completed tasks 1.55 times more often than the VLA-based DYNA-1 .
• In pairwise comparisons, DYNA-2 won 65% of trials versus DYNA-1's 29% .

Aggregate Benchmark Results (14–15 tasks across 3 robot embodiments)

Mean normalized performance rose monotonically with pre-training scale: 20% → 28% → 45% → 53% of the attainable maximum as data scaled from 1k → 10k → 100k → 1M hours . At the 1M-hour scale, DYNA-2 achieved the strongest real-world performance on 9 of 14 tasks . Key task-level improvements included:

• Lockbox Key Turning: from 0% success (up to 100k hours) to 90% success at 1M hours .
• Bottle Cap Untwisting: from 10% success to 40–50% at 1M hours, using only ~10 minutes of robot post-training data .
• Targeted Drink Retrieval (language following): improved from 58% to 83% .

Real-World Manufacturing and Customer Deployment

• High-precision manufacturing tasks: success rates jumped from ~20% to 80–90% purely through increased pre-training scale, with no changes to post-training data .
• Zero-shot deployment at a customer site: DYNA-2 recorded an 87% quality pass rate compared to 46% for DYNA-1 — a 41-point improvement .
• The video co-training algorithm lifted instruction-following scores by 133% on tasks requiring distinct motions in response to user commands .

The Scaling Law Dyna Robotics Observed

Dyna Robotics reported what it calls "the first true scaling law in robotics powered entirely by human data" . Key characteristics:

• Human-to-robot transfer scaling law: As the amount of egocentric human video pre-training increased across four orders of magnitude (1k → 1M hours), robot performance on unseen robot hardware improved predictably, monotonically, and log-linearly — a pattern the company identifies as a scaling law .
• Cross-embodiment transfer: The scaling held across two different robot embodiments (a stationary bimanual platform and a semi-humanoid prototype) that the model had never seen during pre-training .
• Zero-shot action prediction: Scaling human video data also improved the model's ability to predict actions on held-out robot video data that was never shown during training .

The company explicitly states that this is a "first-of-its-kind" demonstration that scaling human observational data, not robot action data, can serve as a reliable axis for improving physical robot intelligence .

Why This Matters

If the claims hold up to independent verification, Dyna Robotics has shown a practical path to scaling robot intelligence without the bottleneck of collecting expensive robot action data. Human video exists at effectively unlimited scale. The company says it is already working toward 10 million hours of training data, which would enable robots to master new physical tasks with just hours of local fine-tuning .