EyeRobot 2.0: active gaze lifts robot success from 52% to 67% — no wrist cameras needed
stepjamUK · x · 2026-10-05
EyeRobot 2.0 from UC Berkeley and Amazon FAR (advised by Ken Goldberg and akanazawa) drops wrist cameras entirely: one stereo head actively moves its gaze to whatever matters at each step — sharp in the middle, blurry at the edges, like human eyes.
Across 7 real tasks (capping a marker, zipping a bag, pulling a pan from a toaster oven), with the same demonstrations, policy architecture and 25 trials per task:
- Fixed stereo camera: 27%
- With wrist cameras: 52%
- Active gaze: 67%
When a grasped tool blocks wrist cameras, that policy drops to fixed-camera levels; gaze keeps working. Gaze is trained with RL, no gaze demonstrations — where it looks is where it reaches.
More from Embodied
- Panasonic building its own humanoid robot, targeting 2029 production in its own factories — CyberRobooo · 2026-10-05
- FailBank Turns Runtime Shield Feedback into Persistent VLA Policy Gains (+25.4 Points Success) — notredame · 2026-10-05
- Reka releases Rho-1 research preview: a 19B omni model unifying text, images, video and robot actions — RekaAILabs · 2026-10-05
- NVIDIA's ENPIRE lets coding agents self-improve robot policies in the real world, hitting 99% success — chris_j_paxton · 2026-10-05
- Sensori demos humanoid Yuri: natural voice chat plus VR and exoskeleton teleoperation — YuXiang_IRVL · 2026-10-05
- UniWAM unifies world model and action policy, hits SOTA on LIBERO and RoboTwin benchmarks — LexiLove · 2026-10-05