RL environments become a business: Mercor at $20B valuation, Lumos launches long-horizon robot RL envs

量子位 · wechat · 2026-11-16

RL training environments are becoming core post-training infrastructure. Mercor raised $750M at a $20B valuation with a 5M-expert network; SurgeAI used its own office-task RL environments to post-train Qwen3.5-122B-A10B (+5.7 pts on SWE-Bench Pro); Handshake AI hit $1B annualized revenue in about a year. On the physical AI side, Lumos Intelligence (光轮智能) launched RoboFinals RL Environments at CoRL 2026: an 8+ minute long-horizon car-repair sim task combining dexterous hands, tool use, and deformable objects, built on its full-stack SimFoundry solver stack with Agent-driven physical calibration, paired with a same-source benchmark for train-eval alignment.

Original post →

More from Embodied

Embodied channel →