Humanoids

EATR-Stereo Lets a Head-Cam Humanoid Keep the Other Eye

Robb Harlan 4 min read

A humanoid’s head cameras bounce with every step. One view gets a hand in the way; the other still sees the bottle. EATR-Stereo, from Harbin Institute of Technology and Honor, is a token-routing add-on that keeps the pretrained primary view and only pulls in the paired stereo frame when the body says it would help.

On a 33-DoF HONOR Omega 1.0 with a 37-D proprioceptive state (33 joints plus a 4-D base quaternion), the paper reports 60.0% full-task success, 100.0% grasp success, and 80.0% stage success on search–approach–grasp–place–return jobs that last more than 100 seconds.

HONOR Omega humanoid with head-mounted stereo cameras and a bottle search-grasp-place task
Physical evaluation on Omega 1.0: stereo head cameras, a 100-second bottle task, and an occlusion test. Source: Wu et al., arXiv:2608.17453.

How the routing works

The authors freeze a Cosmos vision–language model inside GR00T N1.7. Primary-view tokens stay in the original pathway. The other camera becomes Cross-View Auxiliary Tokens (CVATs) by querying the synchronized auxiliary sequence. A body-segmented encoder (legs, arms, head, waist, pose) then gates those tokens token-by-token.

They trained nine configurations on 1,000 segmented-task demonstrations for 60,000 steps. Policies run at 10 Hz on an RTX 4090 and emit 30-step action chunks, stitched with a local cubic B-spline.

Diagram comparing a fixed-camera arm to a biped with head-mounted stereo and paired-view occlusion
Why stereo on a walking humanoid is not the same as a tripod arm camera. Source: Wu et al., arXiv:2608.17453.

What beat the baselines

Over 180 physical trials (20 per method, split in-distribution / unseen bottle positions):

  • EATR-Stereo: 12/20 full-task (60%), 20/20 grasps
  • StereoPolicy on the same backbone: 45% full-task, 85% grasp
  • Default two-image GR00T: 35% full-task
  • Monocular GR00T: 35% full-task, 55% grasp

Under severe asymmetric occlusion — bottle gone from the primary view, still partly visible in the auxiliary — recovery hits 80% (8/10) in 22.4 s mean, versus 30% for CVAT without the body router.

Training cost stays close to frozen GR00T: 10.95 hours versus 10.35 hours on 16 H20 GPUs. An estimated-depth VLA that fine-tunes four VLM layers reached the same 80% stage rate in 41.8 hours.

In RoboCasa365 simulation (18 tasks, 360 trials, no wrist cameras), EATR-Stereo led at 43.33% (156/360).

A Human’s Take

Stereo on a biped is mostly a self-occlusion problem. Hands, torso, and gait keep eating the primary frame. Routing the spare camera instead of fusing it into mush is the part I would steal.

One hundred percent grasp on twenty trials is a lab number, not a warehouse number. I want the same routing on a longer shift, with the bottle behind a moving coworker instead of a staged hand.

Sources