FoveaVLA lets the robot policy choose where to look: 70.8% vs 52.1% for wrist cams
stepjamUK · x · 2026-11-05
Stanford/Princeton researchers built FoveaVLA on π0.5: at each action chunk the policy predicts the bounding box it needs to see next and crops it from the full-resolution external image, instead of squashing 1280x720 to 224x224.
- One external camera + depth, no wrist cam: 34/48 real-world trials (70.8%) vs 25/48 (52.1%) for wrist-equipped π0.5; retraining π0.5 on external view alone drops to 6.3%
- Ablations: crop alone 20.8%, depth alone 6.3%, both 70.8%
- Code and models open; eval is small (12 trials per suite)
More from Embodied
- OpenRoboto Signs First Commercial Robotics Data Pilot at Major European Factory — const_reborn · 2026-11-05
- Dev clones physical objects in AR on PICO's new device, built in hours — stspanho · 2026-11-05
- Neuracore deploys OpenArm policy on open-source robot arm — stepjamUK · 2026-11-05
- ETH Zurich's ADAPT maps text directly to humanoid joint actions with ~2ms diffusion denoising — breadli428 · 2026-11-05
- Hacker turns old Echo Dot into a Hermes agent speaker, open-sources the Home Assistant setup — Teknium · 2026-11-05
- Open-source humanoid BRIDGE: 88cm, 12.5kg, ~$1.5K platform cost — philfung · 2026-11-05