Humanoids

PASSAGE Walks a G1 Through 50 Unseen Clutter Layouts Onboard

Shar Hendrix 4 min read

Stepping over a pallet, ducking a beam, and squeezing past a case are different skills if you train them that way. PASSAGE is a single destination-conditioned planner plus a whole-body tracker that is supposed to pick the mix from onboard geometry. The paper, dated 16 September, puts a Unitree G1 through 50 physical layouts with no prebuilt map and no offboard computer.

PASSAGE system diagram with real G1 photos stepping over, ducking under, and moving sideways through clutter
Data, elevation map, flow planner, tracker, and real G1 stills. Source: Ma, Zeng et al., arXiv:2609.18732, teaser.

100 hours in a headset, 1,500 fake corridors

Operators in a Noitom PN Link mocap suit and a VR headset walk 10 m procedural corridors stuffed with ground, side, and overhead blocks. Collisions get haptic buzz and the take is thrown out. After retargeting to a 29-DoF G1 and screening, they keep 19,310 sequences, about 100 hours, across 1,500 scenes. Validated obstacle rescales (0.5–1.5 scale, ±15°) expand that to roughly 1,000 hours of motion–scene pairs without more capture.

A conditional flow-matching planner emits 0.5 s (25-frame) references at 6.25 Hz from four history frames, a local destination, and a torso-centered three-layer elevation map (support, lateral block, overhead clearance). Real-time chunking blends the previous plan into the next so the robot does not twitch at chunk boundaries. A ScaleBFM tracker, perception-augmented, runs at 50 Hz. Then they freeze the tracker and RL-fine-tune only the planner on closed-loop rollouts.

Onboard: Manifold Tech Odin LiDAR, ROG-Map occupancy, TensorRT FP16 planner on a Jetson AGX Orin, CPU tracker, joint commands at 500 Hz.

Operator in a mocap suit stepping through a VR clutter scene with an egocentric inset
VR-guided capture. Source: arXiv:2609.18732, Figure 3.

Sim scale, then 50 real layouts

In MuJoCo, 150 held-out scenes × 5 rollouts. Across three seeds, scaling captured data from 6 to 100 h lifts mean success 84.3% → 96.4% and contact-free success 48.1% → 68.9%. The final augmented checkpoint hits 98.7% success and 70.3% contact-free. CAT’s released generalist, zero-shot on the same test distribution, is 70.3% / 14.0%. Ablations are brutal: drop RTC and contact-free success falls to 24.5%; drop planner RL, 26.4%; swap in the base tracker, 18.4%.

Hardware: 50 distinct layouts, destination-only commands, one attempt each. 50/50 reach the finish without a fall or safety stop. 45/50 (90%) are contact-free. The five contacts are the remaining problem, not falling over.

Simulated G1 under a block arch beside onboard LiDAR reconstruction of a real arch and pallet
Same three-layer geometry in sim and from live LiDAR. Source: arXiv:2609.18732, Figure 4.

Project page: galaxygeneralrobotics.github.io/PASSAGE.

A Human’s Take

Fifty layouts with one planner and a backpack Jetson is the demo I want more of. Contact-free at 45/50 is the honest number: it finishes the course, and sometimes it still kisses the foam. If they add stairs and thin cables like the conclusion teases, keep the same 50-layout rule. I will be watching the contact column, not the highlight reel.

Sources