966 real navigation episodes, 24km: ECCV study reveals which visual encoder components matter for robots
chriswolfvision · x · 2026-09-12
A companion post to the ECCV paper (arXiv:2606.21216) by Steeve Janny, Leonid Antsfeld and Christian Wolf: 966 static point-goal navigation episodes over 24km in a real building, evaluating state-of-the-art visual encoders under realistic conditions. Key findings: encoder power stems largely from CV pre-training losses; heterogeneous multi-teacher distillation yields complementary skills; principled spatial bottlenecks produce interpretable affordance-linked features; and pre-training policies on privileged Lidar input before fine-tuning beats RGB-only training.
More from Embodied
- She hacked her doorbell camera to rank Halloween costumes — and open-sourced it — jeff_weinstein · 2026-10-31
- Surgical robot sim benchmark: only 2 of 10 frontier models completed task, all 213 success claims wrong — Exp_Mark · 2026-10-31
- China adds embodied AI robot technician and digital twin engineer to official jobs list — nordicinst · 2026-10-31
- YC-backed Axonnic launches EMG sleeve to capture human hand skills for robot training — ycombinator · 2026-10-31
- VITURE launches $599 Phantom Beast XR glasses in Phantom Blade Zero collab — Scobleizer · 2026-10-31
- Agility Robotics lost $138.1M in 2025 as Brett Adcock's doom prediction nears — Rewkang · 2026-10-31