Robocurve independent eval: Dream-1 tops robot control at 16x lower cost
Benchmarking outfit Robocurve independently evaluated Dream Labs' Dream-1 before its release, concluding that the model sets a new Pareto frontier for robot control, and subsequently released all raw run data. If the results hold up under replication, they could materially reshape the cost and performance landscape for robotics foundation models.
Confirmed
- Robocurve reports Dream-1 is 16x cheaper and 1.7x faster than the previous best approach.
- On the RoboDojo-RC Tier 1 real-robot leaderboard, Dream-1 took first place with 59.3%, becoming the first model to break 50% and far ahead of other evaluated models including the Gemini family (the post mentions Gemini 4 as a comparison).
- Cost per run is about $0.15.
- Robocurve published all 120 Dream-1 run records, including full conversation transcripts, logs, and videos, plus a filterable run browser and CSV downloads.
- The site also opened its full evaluation database of 2,469 runs, filterable by benchmark (RoboDojo-RC Tier 1, RoboHarm, etc.).
Why it matters
- The evaluation took place before the model's release — a third-party independent assessment rather than vendor marketing, giving it higher credibility.
- An order-of-magnitude cost drop plus a speed gain, if the data holds, would significantly lower the deployment barrier for real-world robot tasks.
- Releasing all raw data (transcripts, logs, videos, CSV) enables community replication and auditing, setting a transparency benchmark for robotics foundation model evaluation.
2026-10-20 ~ 2026-10-20 · 5 related posts
- Dream-1 claims new Pareto frontier for robot control: 16x cheaper, 1.7x faster, 59 vs 48 on RoboDojo — scott_e_reed · 2026-10-20
- Robocurve releases all 120 Dream-1 robot runs with transcripts and videos — santiviquez · 2026-10-20
- Robocurve releases all 120 Dream-1 robot runs with full transcripts, videos and leaderboard — gharik · 2026-10-20
- Robocurve real-world evals: Dream-1 tops robotics at 59%, costs $0.15 per attempt — gharik · 2026-10-20
1 near-duplicate retellings: gharik