의사결정 지능 및 학습

We build mathematically grounded yet practical reinforcement learning systems: agents that learn reliable, robust behavior from imperfect, heterogeneous, and large-scale data, and adapt efficiently to new objectives, rewards, and constraints. Principled algorithms with provable guarantees for data-efficient, reliable decision making. Learning reliable policies from fixed, previously collected data, without costly or risky online interaction. Optimizing behavior under safety and cost constraints and multiple competing objectives. Robust behavior from demonstrations, observations, and noisy or biased supervision, including reliable off-policy evaluation. Exploiting the structure of decision problems (symmetries, embodiment, skills, and multiple agents) to generalize across tasks and environments. Toward generalist agents pretrained across tasks that transfer to new objectives, rewards, and environments.