ROOF teaches a humanoid to make room beneath a low obstacle and then stand, using one learned recovery policy rather than separate escape and get-up routines. The policy uses obstacle geometry during training but relies only on the robot’s own body feedback at deployment, and the researchers tested it in simulation and across seven real-world scenes on a Unitree G1.
What did the researchers build?
ROOF is a learned behavior for humanoids that fall where there is not enough headroom to stand up normally. A robot under a table, shelf, or fallen panel cannot simply raise its torso: it may hit the obstruction before it can get upright. It first needs to create clearance while keeping, or finding, enough support to rise afterward.
The system learns that entire recovery as one connected task. It can move its own body out from under an obstacle, push a movable obstacle, or combine those actions. Rather than prescribing a particular escape direction or requiring a set sequence of contacts, the training reward guides the robot toward making more room while staying in a useful supported posture.
The researchers tested the approach in simulation and on a Unitree G1 humanoid. Their simulated setup includes different kinds of obstacles, while the hardware tests use furniture and arrangements of multiple objects. A key design choice is that the obstacle’s geometry informs training, but the deployed policy uses the robot’s proprioception: feedback about its own body and joint state. It does not need obstacle geometry as an input while carrying out a recovery.
What results did ROOF achieve?
The evaluation covers two simulated obstacle types and seven real-world scenes. In simulation, one plate is constrained to move vertically: it can be lifted but cannot slide or tilt, so the robot needs to reposition itself to get clear. Another plate is free to move, allowing the robot to push, slide, or tip it aside. An unobstructed-ground condition provides a control case for recovery without an overhead obstacle.
On hardware, the Unitree G1 demonstrates recovery beneath furniture and in multi-object arrangements that were absent from training. This tests whether the learned behavior can handle scene layouts beyond the specific training instances, rather than only repeating a familiar setup.
The reported results show that one policy can coordinate clearance creation with standing, including scenarios where moving the robot and moving the obstacle are both possible. The supplied paper text does not include success rates, recovery times, or numerical comparisons against other methods, so it does not support a quantitative ranking. The concrete evidence here is the range of simulated obstacle behaviors and the seven hardware scenes.
How does ROOF work?
ROOF’s central training signal measures whether the robot is making geometric progress against an obstruction. The reward tracks overlap between the robot and obstacle, as well as separation between them. In plain terms, it encourages the robot to reduce the part of its body that is blocked and create enough distance to continue recovering.
Crucially, that signal does not tell the robot exactly how to move. A fixed instruction such as “always slide left” or “push with the arm” would be brittle when obstacle position, shape, or mobility changes. Instead, ROOF rewards progress while the robot stays low and supported, leaving the learned policy to choose an escape direction and useful contacts. Depending on the situation, it can move itself, displace an obstacle, or do both.
The policy also uses a learned human-motion prior: a model that encourages coordinated movement rather than awkward, disconnected motions. Posture-dependent recovery objectives help shape the transition from supported repositioning to a controlled rise. The researchers additionally use foot-motion and joint-stall costs during refinement to discourage undesirable motion and help improve transfer to hardware. The motion prior guides training but is discarded at deployment.
The overall sequence is therefore learned as a coordinated behavior, not a hand-coded checklist. Obstacle geometry supplies guidance during training; the deployed policy acts from proprioceptive feedback. That division lets the training process account for the task’s spatial constraints without requiring the robot to receive an explicit geometric map when it recovers.
Why does obstacle-aware recovery matter for robotics?
A robot that can stand up only when it has open space around its torso is vulnerable in cluttered workplaces and homes. Low furniture, shelving, stored materials, and objects displaced by an accident can turn an ordinary fall into a recovery problem that a standard get-up motion cannot solve. ROOF addresses the missing step: make space first, while preserving a path to stable standing.
The method also points toward a practical way to train contact-rich behavior. Contact with the environment is not always a failure; a robot may need to brace, slide, or push. ROOF’s reward guides the outcome—less obstruction and useful support—without requiring engineers to specify every contact and motion in advance.
For buyers and integrators comparing platforms, recovery is one part of evaluating whether a humanoid can operate around everyday obstacles. Browse humanoid robots on Robot Overflow to explore the broader category. ROOF is a research demonstration, not evidence that every commercial humanoid can recover this way, but it highlights a concrete capability to look for in systems intended to work in tight spaces.
What limitations and open questions remain?
The results cover a particular robot, a bounded set of simulated obstacle types, and seven hardware scenes. They establish that the approach can work in those settings, but the supplied text does not report numerical success rates, recovery times, or controlled comparisons with alternative methods. That makes it difficult to judge how often ROOF succeeds or how much it improves on existing recovery policies.
The method also relies on obstacle geometry during training, even though deployment uses proprioception alone. More evidence is needed about performance with unfamiliar obstacle shapes, unstable or delicate objects, changing surfaces, and falls that leave the robot in very different postures. The paper’s described tests do not establish how reliably the robot can detect when an attempted push is unsafe or when recovery should be abandoned.
What should readers know about ROOF?
Does ROOF require a map of the obstacle at deployment? No. Obstacle geometry guides training, while the deployed policy uses the robot’s proprioception.
Can the robot move the obstacle as well as itself? Yes. The learned behavior can move the robot, push a movable obstacle, or combine both actions.
Was ROOF tested on a physical humanoid? Yes. The researchers evaluated it on a Unitree G1 across seven real-world scenes.
Are numerical success rates or recovery times reported in the supplied text? No. The available text describes the evaluation and demonstrations but provides no such performance figures.
What should readers take away?
ROOF treats getting up beneath an obstacle as a combined problem of creating clearance and maintaining support, rather than applying a standard stand-up motion. Its hardware demonstrations show a promising way to train humanoids for recovery in clutter, while leaving important questions about reliability and broader deployment open.