SONG-Bench Dynamic 3DGS Social Navigation Episodes
SONG论文描述的动态社会导航仿真资产与基准：1,000个SAGE-3D照片级3DGS环境、500个HumanGS动态人像，LLM语义意图、Recast/DetourCrowd轨迹和Kimodo动作生成500条评测episode，并按100/200/200分为困难/中等/简单。论文称官方episode以确定性缓存和脚本发布，但没有给出官方代码、数据入口、文件清单或聚合许可证，当前不能计作已开放。
- 开放状态
- 项目已核验·制品未开放
- 许可证
- No aggregate dataset or asset license published; upstream SAGE-3D and HumanGS terms cannot be assumed to cover SONG-Bench
- 商业使用
- not established
- 人工评审结论
- paper_verified_3dgs_social_navigation_benchmark_not_released
- 机器人
- Unitree Go2 physical evaluation、vision-based point-goal navigation agents
- 任务
- vision-based point-goal social navigation around one or more moving pedestrians
- 模态
- rendered RGB-D、3DGS scene assets、dynamic 3DGS human avatars、semantic intent text、pedestrian trajectories、human motions、goal poses、collision and distance metrics
- 格式
- paper says deterministic episode caches plus scripts and YAML/JSON, but no files were located
- 采集/生成方式
- custom Gaussian-splatting social-navigation runtime with Recast/DetourCrowd trajectory generation
- 已知问题
- No official artifact URL, manifest, checksum, dataset license or reproducible asset build. Dynamic pedestrians do not react to the robot, benchmark comes from the authors, and physical success is only 5-20% with no independent reproduction.