RescueBench

RescueBench is a mission-level benchmark for evaluating embodied agents in simulated search-and-rescue (SAR) environments. Given multimodal clues and egocentric observations, an agent must autonomously explore an unfamiliar environment, locate and approach the target, return to the rescue area using spatial memory, and complete the final handoff.

Each mission is decomposed into four sequential stages:

1. Explore and locate the target.
2. Approach and retrieve the target.
3. Return to the rescue area using observation history and spatial memory.
4. Complete the handoff and finish the mission.

This repository hosts the RescueBench evaluation data and representative demonstrations across different environments, agents, and difficulty settings.

Human-Control Demonstrations

HongKongStreet / Human / L4

Tokyo / Human / L3

DowntownWest / Human / L2

These GIFs are representative examples illustrating the human-control interface and complete mission workflow; they are not the full set of participant trajectories.

Benchmark Demonstrations

FlexibleRoom / ROCKET	FlexibleRoom / ROCKET	FlexibleRoom / YOLO Planner
DowntownWest / Uni-NaVid	Forglar Map / ViNT	Forglar Map / Uni-NaVid
Tokyo / Uni-NaVid	HongKongStreet / Uni-NaVid	SuburbNeighborhood Day / ROCKET

Heterogeneous Multi-Agent Cooperation

Third-person view

UAV view

Ground-agent view