Jisu Han

Benchmarks whether foundation-model robot policies adjust low-level navigation to safety context. It uses matched scenes where obstacle semantics or safety demand changes, exposing failures like treating a baby too similarly to a rigid object.

Introduces option-aware, temporally abstracted value learning for offline goal-conditioned RL. By reasoning over options instead of only primitive actions, the agent can make more reliable long-horizon goal-reaching decisions from static datasets.

Proposes a hierarchical, modular network for non-prehensile manipulation such as pushing. The structure separates reusable manipulation components, helping policies generalize across broader object configurations, layouts, and environments.

Reduces visual computation for robot manipulation by adaptively merging and pruning object tokens. The policy keeps task-relevant object information while removing redundant regions, improving efficiency without losing key grounding cues.

Learns a preference model that scores candidate outcomes during online planning in continuous POMDPs. These preferences guide tree search toward actions that better match expert intent, improving decision-making under uncertainty.