Galbot's PASSAGE teaches humanoids to duck, step over and squeeze through clutter

In short: Galbot highlighted PASSAGE on Oct 13, 2026 — a perception-conditioned planner–tracker that lets a humanoid choose how to cross clutter: step over, duck under or turn sideways. Trained on 100 hours of scene-aligned human motion across 1,500 scenes, it runs fully onboard a Unitree G1 with Jetson AGX Orin and lidar, no prebuilt map.

The numbers that matter

In 750 simulated trials on held-out scenes, destination reach was 98.67%; requiring no obstacle contact dropped that to 70.3%. On 50 unseen physical layouts, the team reports all 50 finished without falls or safety stops, and 45 without contact. That is a limited research eval — not a product claim for Galbot's own ET1.

Why it matters

Feet-clear paths still crash shoulders into doorways. PASSAGE's single whole-body control pair composes behaviors without one policy per obstacle type — useful for homes and factories built for people, while the contact gap shows how hard clutter still is.