SIS-Bench: New Embodied AI Benchmark for UAVs
Zhishan Zou · hf · 2026-07-17
Introduces SIS-Bench, a new benchmark for UAV embodied AI designed to evaluate spatial intelligence across two main tracks: spatial understanding and self-awareness.
Benchmark Design
- Unified under a self-in-space setting, examining both space and self dimensions.
- Tasks are organized across three levels: perception / memory / reasoning.
- Data is sourced from 1,646 real-world UAV videos, task-structured and expert-verified.
- Contains 4,856 QA pairs across 13 tasks.
Key Findings
- Existing MLLMs show inadequate modeling capabilities for dynamic, agent-centric processes in UAV scenarios.
- Models exhibit significant imbalances between spatial cognition and self-awareness.
- Performance degrades progressively as the cognitive level moves from perception to memory to reasoning.
Method & Effects
The authors further introduce motion-aware representation, fusing optical flow with visual features into self-relevant dynamic modeling.
- Results show that explicitly modeling motion information consistently boosts perception and memory performance.
- This improvement is evident in both spatial cognition and self-awareness tasks.
- The representation also generalizes well to downstream UAV decision-making tasks.
Conclusion: Advancing UAV embodied spatial intelligence requires not only environmental understanding but also explicit modeling of the "self's" dynamic state in space.
More from Embodied
- Orlando robotaxi ride goes unsupervised in a Model Y — aelluswamy · 2026-07-21
- Anthropic and Physical Intelligence held early M&A talks this spring — steph_palazzolo · 2026-07-21
- EPO improves 3D foundation models by aligning edges, poses, and depth — ducha_aiki · 2026-07-21
- Applied Intuition launches Dana for physical AI, alongside a16z interview on one billion machines — a16z · 2026-07-21
- Hearing Aids Embrace AI: Auracast and OmegaAI Reshape Personal Audio Broadcasting — BrandonSawalich · 2026-07-21
- Argus improves indoor panoramic 3D reconstruction with covisibility and geometry transformers — ducha_aiki · 2026-07-21