Robocurve real-world evals: Dream-1 tops robotics at 59%, costs $0.15 per attempt
gharik · x · 2026-10-20
Robocurve, an independent Public Benefit Corporation, released real-world evaluations of frontier AI on robots. On RoboDojo-RC Tier 1, Dream Labs' Dream-1 scores 59.3% — the first model above 50% — beating Gemini 4 Argon (48.2%), GPT-6.1 Sol (41.7%), Opus 5.5 (38.2%) and other frontier LLMs.
Key numbers:
- Dream-1 costs just $0.15 per attempt (hosted inference) and sits alone on the performance-cost Pareto frontier; GPT-6.1 Sol runs $2.15 and Opus 5.5 high $2.23 per attempt.
- VLA baselines GR00T N1.7 (9.5%) and MolmoAct2 (18.9%) are cheap ($0.05-0.07) but far weaker.
- On RoboHarm v1 harmful-instruction refusal, Opus 5.5 and GPT-6 Sol scored 0, while GPT-6 Astra topped out at 32%.
- Each config ran 120 attempts, capped at 40 model calls (LLMs) or 3,600 control steps (VLAs).
Related event: Robocurve independent eval: Dream-1 tops robot control at 16x lower cost(5 posts)→
More from Embodied
- Tactile Curiosity Drives Robot Grasping Without Task Rewards, CoRL 2026 Paper — pabbeel · 2026-10-20
- Dream-1 tops RoboDojo-RC Tier 1 with score of 59, pushing back on 'LLMs will solve robotics' — scott_e_reed · 2026-10-20
- Robotics practitioner: LLM-generated 'robotics startups' all chase the same fake problems — kscottz · 2026-10-20
- Realtime Local AI Video Restyling: 720p 60fps on a Single RTX 5070 Ti — Fit_Revenue1928 · 2026-10-20
- GPT-6 Astra in spatial XR: Jarvis-style live Earth demo on a PICO device — Scobleizer · 2026-10-20
- Long queues at RTX Spark SF launch for local Qwen demos built with ODS software — TheZachMueller · 2026-10-20