Sim data can't cover the humanoid long tail: the sim2real debate, mapped
herbiebradley · x · 2026-10-16
Herbie Bradley pushes back on "simulation solves the data problem" in a reply, unpacking the embodied-AI data debate:
- Simulating data for e.g. a police humanoid robot would require approximating the full distribution of real-world events to cover the long tail, hitting weather-simulation computational complexity; structured settings like manufacturing are feasible. "Just saying 'you can sim' is far too low-resolution."
- He splits believers into two camps: category 1 thinks models stay data-bottlenecked and sim is a brute-force patch for the long tail; category 2 thinks models will generalize like humans and never need distribution coverage at all.
A substantive technical argument on sim2real and scaling routes for robotics.
More from AGI Musings
- "An LLM doesn't understand anything" gets the perfect parody reversal: neither do you — VoidStateKate · 2026-10-16
- Nina Schick: Europe's Policy Circle Still Doesn't Believe FSD, Orbital Compute Are Real — NinaDSchick · 2026-10-16
- Meta's Superintelligence Lab Runs 200+ Direct Reports—A Lesson for Flat Multi-Agent Systems — peterjliu · 2026-10-16
- Andy Masley lays out his version of effective altruism, pluralism over utilitarian math — RobbWiller · 2026-10-16
- We're forgetting how much of a miracle GPT-3.5-level intelligence was, and it may cost us — EigenGender · 2026-10-16
- Ex-Anthropic's McDonagh Builds Compute Intelligence Agency to Study Frontier AI Costs — McDonaghMatthew · 2026-10-16