Lifelong Visual Navigation on GOAT-Bench 1/10-scale subset (val-unseen)
72.8Success RateOurs
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OursBackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration2026.02 | 72.8 | 56.1 | |
| 3D-MemBackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration2026.02 | 65.1 | 49.3 | |
| CG w/ Frontier SnapshotsBackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration, Method Variant=with Frontier Snapshots2026.02 | 61.5 | 45.3 | |
| ReEXploreBackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration2026.02 | 59.8 | 42.5 | |
| 3D-Mem w/o MemoryBackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration, Method Variant=without Memory2026.02 | 58.6 | 38.5 | |
| Explore-EQABackbone=GPT-4o, Exploration Strategy=MLLM-based Exploration2026.02 | 55 | 37.9 | |
| OursBackbone=Qwen2.5-VL-7B-Instruct, Exploration Strategy=MLLM-based Exploration2026.02 | 53.9 | 31.7 | |
| ReEXploreBackbone=Qwen2.5-VL-7B-Instruct, Exploration Strategy=MLLM-based Exploration2026.02 | 53.2 | 32.6 | |
| 3D-MemBackbone=Qwen2.5-VL-7B-Instruct, Exploration Strategy=MLLM-based Exploration2026.02 | 49.4 | 20.7 | |
| SenseAct-NN Skill ChainExploration Strategy=Traditional Methods2026.02 | 29.5 | 11.3 | |
| Modular GOATExploration Strategy=Traditional Methods2026.02 | 24.9 | 17.2 | |
| Modular CLIP on WheelsExploration Strategy=Traditional Methods2026.02 | 16.1 | 10.4 | |
| SenseAct-NN MonolithicExploration Strategy=Traditional Methods2026.02 | 12.3 | 6.8 |