Galbot’s PASSAGE Helps Humanoids Step Over, Duck Under and Squeeze Past Obstacles

Galbot has highlighted PASSAGE, a research framework that lets a humanoid choose how to move through clutter: stepping over low obstacles, ducking beneath overhead barriers and turning sideways through narrow gaps. The company shared the work in an October 13 post, following a September research paper.

Key Takeaways

Show all

• PASSAGE combines environmental perception, motion planning and whole-body control to negotiate obstacles.
• The team collected 100 hours of human motion across 1,500 virtual scenes.
• Its reported 98.67% simulation success rate measures reaching the destination; contact-free success was 70.3%.

The practical challenge is deciding which parts of a robot’s body must move to fit through a space. A clear route for its feet may still leave its shoulders or head on a collision course.

According to the project page, PASSAGE pairs a planner that generates short movement sequences with a controller that executes them while responding to the surrounding geometry. One planner–controller pair selects and combines movements without separate policies for each obstacle type.

The training data connects human movement with the environment in which it occurs. Operators used virtual reality and inertial motion capture to demonstrate traversal. On the robot, LiDAR supplies environmental information, and the system runs on a Jetson AGX Orin without prebuilt maps or offboard computation.

The paper gives useful context for Galbot’s headline result. Across 750 simulated trials in 150 held-out scenes, the final model reached the destination in 98.67% of attempts. Requiring completion without touching an obstacle reduced that figure to 70.3%.

Physical testing used a Unitree G1 across 50 previously unseen layouts, each attempted once. The authors report completion without a fall or safety intervention in all 50 trials, with 45 completed without obstacle contact. These are the research team’s results from a limited evaluation, rather than evidence of deployment reliability.

The G1 is a research platform here, distinct from Galbot’s own ET1, whose three versions start at a reported 79,000 yuan. PASSAGE’s published evaluation does not establish the same capabilities on ET1 or explain the choice of hardware.

The work adds to Galbot’s broader research into humanoid movement. We previously covered RoboGesture’s real-time gestures coordinated with speech, also demonstrated on a Unitree G1, and AstraTennis’s live tennis demonstration. PASSAGE addresses another part of that challenge: adapting whole-body movement to the geometry of everyday spaces.

For humanoids intended to operate in furnished rooms and crowded workplaces, adapting body posture to available space is a useful capability. PASSAGE offers a concrete demonstration of that coordination, while the contact figures show where further improvement is needed. The project page currently lists code as coming soon.