Galbot's PASSAGE Lets Humanoids Pick How to Clear Obstacles

A humanoid robot that can move through a cluttered room has to solve a problem that sounds trivial and is not: deciding which parts of its own body need to move to fit through a space, since a clear path for its feet can still leave its shoulders or head on a collision course. Galbot, the Beijing-based embodied AI company founded in May 2023 by Peking University researcher He Wang, shared a framework called PASSAGE on October 13, 2026, that lets a humanoid choose in real time whether to step over a low obstacle, duck beneath an overhead barrier, or turn sideways to squeeze through a narrow gap, following a research paper the company published in September.

Galbot builds the ET1, a bipedal humanoid aimed at entertainment and performance applications that it prices from 79,000 yuan, approximately US$11,000, across three hardware versions, a price point considerably below many Western humanoid platforms and a reminder that Galbot's commercial strategy runs through volume and accessible pricing rather than premium industrial contracts alone. PASSAGE itself was demonstrated on a different machine, a Unitree G1 serving as a research platform, which matters for how a buyer should read the results: Galbot is publishing a capability result, not a product claim about what ships on the ET1 today.

The Numbers Behind a Demo That Looks Simpler Than It Is

PASSAGE works by joining two parts: a planner that works out short bursts of movement, and a controller that carries them out while adjusting on the fly to whatever geometry surrounds the robot, all inside one combined system rather than a separate policy for every type of obstacle a robot might encounter. According to Galbot's project page, the training data connects human movement directly to the environment it occurs in: operators used virtual reality and inertial motion capture to demonstrate traversal, and the team collected 100 hours of this human motion data across 1,500 virtual scenes. On the robot itself, LiDAR supplies environmental information and the whole system runs on an Nvidia Jetson AGX Orin onboard computer, without prebuilt maps or offboard computation, meaning the robot is reading and reacting to a space it has never mapped in advance.

The headline result, a 98.67 percent success rate across 750 simulated trials in 150 held-out scenes, measures only whether the robot reached its destination. Requiring that it do so without touching an obstacle along the way cuts that figure to 70.3 percent, a gap that is the single most important number in this release for anyone evaluating the technology rather than the press account of it. A robot that reaches a goal by brushing past a shelf or grazing a doorway is acceptable in a research benchmark and potentially costly in a warehouse, hospital corridor or retail floor where the obstacles are fragile, expensive or occupied by people.

Physical Testing Tells a More Conservative Story Than Simulation

Galbot also ran PASSAGE on a physical Unitree G1 across 50 previously unseen layouts, each attempted once rather than averaged over repeated tries, which is a meaningfully stricter test than the simulated trials. The robot completed all 50 attempts without a fall or a safety intervention, and 45 of those 50, or 90 percent, completed without any obstacle contact at all. That physical-world contact-free rate is notably higher than the simulated 70.3 percent figure, though the far smaller sample size, 50 real attempts against 750 simulated ones, means the physical result should be read as an encouraging early signal rather than a statistically robust claim on its own. Galbot's own published framing is appropriately restrained here: the authors name collision avoidance, not course completion, as the capability still most in need of improvement, an admission that is more useful to a buyer than a cleaner-sounding marketing claim would have been.

Galbot has built a pattern of publishing these capability demonstrations on the G1 rather than its own hardware, including RoboGesture, a system for coordinating real-time gestures with speech, and AstraTennis, in which a Galbot-trained humanoid completed more than 100 consecutive rallies in a live tennis match. PASSAGE extends that research line into whole-body obstacle negotiation, which, combined with gesture coordination and the fine motor control a tennis rally demands, starts to sketch a capability stack aimed squarely at humanoids operating in the kind of unstructured, human-occupied spaces where entertainment and service robots, Galbot's stated commercial focus, actually have to work.

Why Whole-Body Obstacle Negotiation Is Harder Than It Looks From a Spec Sheet

Most published humanoid locomotion results, from gait stability to stair climbing, treat the robot's path as a known or mostly predictable variable: the robot is told where to go, and the research question is whether its legs and balance control can get it there. PASSAGE addresses a different and arguably harder problem, deciding what the body should do when the destination is clear but the direct path is not, a decision that has to account for the robot's shoulders, head and arms as much as its feet. A warehouse aisle narrowed by a parked cart, a doorway with an overhead pipe, a hospital corridor with a gurney parked at an angle, all of these are routine in human-occupied spaces and all of them require a robot to reason about its own body dimensions against a geometry it has not seen before, rather than simply following a pre-mapped route.

That is why the planner-controller architecture Galbot describes matters more than it might appear from the project page's brief summary. A system built around separate policies for separate obstacle types, stepping over, ducking, turning sideways, would need to classify every new obstacle into one of those categories before acting, a brittle approach once a real environment presents something that is partly all three at once, like a low shelf jutting from a narrow gap. A single combined planner and controller that selects and blends these movements without that classification step is a more generalizable design, and the 98.67 percent destination-reaching rate across held-out, previously unseen scenes is the evidence that the generalization claim holds up at least in simulation, even if the contact-free rate shows there is real room left to tighten the controller's precision.

What a Procurement Team Should Take From a Research Platform Result

The most useful fact for a buyer evaluating humanoid navigation claims broadly is the one Galbot did not have to disclose but did: PASSAGE's published evaluation does not establish the same capabilities on Galbot's own ET1, and the company offers no explanation in its materials for why the G1 was chosen as the demonstration platform over its own hardware. That is not necessarily a red flag, research teams frequently validate a software capability on whichever robot is best instrumented for the experiment, but it does mean a buyer interested in deploying obstacle-aware navigation on an actual ET1 unit should treat this release as evidence that Galbot's research division can produce the capability, not as confirmation that the capability ships on the product Galbot sells today.

For procurement teams evaluating humanoid platforms for warehouses, hospitals or crowded commercial floors, whole-body obstacle negotiation is exactly the kind of capability that separates a robot that works in a controlled demo area from one that can be trusted in a furnished, unpredictable space shared with people. PASSAGE's own numbers, a 90 percent contact-free rate across a genuinely blind physical test, are a meaningfully strong early result for that use case, and the gap between that figure and a deployment-ready guarantee is exactly the kind of evidence a serious buyer should ask Galbot, or any humanoid vendor making similar claims, to close before a purchase order rather than after one.

It is also worth situating PASSAGE against the broader field of humanoid navigation research coming out of both Chinese and Western labs this year. Several competing approaches have pursued obstacle traversal through simpler, single-purpose policies, such as climbing-specific controllers tuned narrowly for stairs or platforms, trading generality for a higher success rate within their narrower domain. Galbot's bet with PASSAGE is the opposite: a single system designed to generalize across obstacle types at some near-term cost to precision, on the premise that a real deployment environment will present combinations of obstacles no narrow controller was trained for. Whether that bet pays off commercially depends on how quickly Galbot and its research peers can close the gap between the 98.67 percent destination-reaching rate and the 70.3 percent contact-free rate, since a generalist system that still clips obstacles on three attempts in ten is not yet a system a risk-averse buyer can insure against property damage in an occupied facility.

The code behind PASSAGE is listed as coming soon on Galbot's project page rather than available today, which means outside researchers cannot yet independently verify the reported numbers by rerunning the experiments themselves. Independent replication has become an increasingly important credibility test in embodied AI research generally, given how often early capability claims from both Chinese and Western humanoid labs have proven narrower in practice than their initial framing suggested. Procurement teams with the technical staff to evaluate a released codebase should treat that eventual release, rather than this announcement, as the point at which PASSAGE's claims become fully auditable.

This analysis synthesizes company statements and published research findings as of the publication date and should not be read as investment, financial, or professional advice; it is provided for general information purposes only.