Princeton's RAVEN stores raw visual embeddings as robot memory, beating caption-based baselines
shahdhruv_ · x · 2026-11-03
Researchers from Princeton and UPenn present RAVEN (RetrievAl via Visual Embeddings for Navigation), an agentic memory system for long-horizon robotic QA and navigation, accepted to the SISREN Workshop at CoRL2026.
- Core idea: typical robot memories caption everything into text first, losing visual detail; RAVEN stores visual embeddings directly, sidestepping the lossy image-to-text bottleneck.
- How it works: every memory is indexed with spatial pose and world-clock timestamp in a vector database, and a VLM agent combines semantic, temporal, spatial, and image retrieval tools grounded in a spatial map.
- Evaluation: consistently outperforms caption-based baselines on NaVQA, FindingDory, and a new RAVEN-QA benchmark spanning web, simulated, and real-world videos.
- Deployment: demonstrated on a Unitree Go1 navigating real indoor environments.
More from Embodied
- Google's Suncatcher satellite runs Gemma 270mi n space—power limits mean only one TPU lit at a time — demian_ai · 2026-11-03
- Tutor Intelligence Puts Hourly-Billed AI Robots in US Factories — a16z · 2026-11-03
- Engineering World Model SKY-001 demos designing a Jaguar car with thousands of complex parts — Scobleizer · 2026-11-03
- PhysiCalWorld Calibrates World Models to Scene Physics with Label-Free In-Context Interaction Clips — Scobleizer · 2026-11-03
- Augmentic debuts photoreal robot simulation, rendering Isaac Sim G-buffers into realistic frames — Scobleizer · 2026-11-03
- YC-backed Darwin Robotics debuts general-purpose robot technicians for data centers — ycombinator · 2026-11-03