ROOF: Learning Humanoid Fall Recovery under Overhead Constraints
ROOF: Learning Humanoid Fall Recovery under Overhead Constraints
This paper introduces ROOF, a reinforcement learning method that enables humanoids to recover from falls while navigating overhead obstacles using only proprioception. The approach combines contact-supported geometric escape guidance with learned motion priors and posture-dependent support costs. In simulation, the method achieves an 88.5% recovery rate under liftable plates compared to 18.9% without guidance, and a 98.6% rate under free plates with randomized properties. On a Unitree G1 robot, the policy successfully recovered in 95 out of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements not present during training.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
attempts
95
Reported trials
View original evidence
18.9% without geometric guidance, and in 98.6% of trials beneath free plates with randomized size, mass and placement. On a Unitree G1, the deployed policy recovered in 95 of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements absent from training. Project page: https://lu-yidan.github.io/ROOF/
Open source S4attempts
150 trials
Reported trials
View original evidence
18.9% without geometric guidance, and in 98.6% of trials beneath free plates with randomized size, mass and placement. On a Unitree G1, the deployed policy recovered in 95 of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements absent from training. Project page: https://lu-yidan.github.io/ROOF/
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
tacle overlap without prescribing an escape path. To achieve this, a learned motion prior, together with posture-dependent support and joint-stall costs, is proposed to shape a single policy that learns when to clear the obstacle, how to establish support, and when to complete the rise, all using proprioception alone. In simulation, it recovers in 88.5% of trials beneath plates that can only be lifted, compared with
Open source S3
18.9% without geometric guidance, and in 98.6% of trials beneath free plates with randomized size, mass and placement. On a Unitree G1, the deployed policy recovered in 95 of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements absent from training. Project page: https://lu-yidan.github.io/ROOF/
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org