Researchers have trained a 29-joint Unitree G1 humanoid to absorb a learned safety boundary into its control policy before deployment. Their LIMBO framework produced dodgeball-avoidance and low-obstacle behaviors on real hardware without running the learned safety critic or an additional safety filter on the robot.
The work addresses a scaling problem in safety-critical robot control. A control barrier function defines which actions keep a system away from failure, but writing one for a whole humanoid can require an accurate dynamics model and hand-designed constraints for tightly coupled motion. LIMBO instead learns a state-action control barrier function, or Q-CBF, from examples of actions and their resulting states, together with a simpler description of failure.
Training proceeds in two stages. First, LIMBO estimates which actions remain recoverable and concentrates simulated experience near the boundary where safety is most uncertain. Second, that learned critic acts as a teacher: it scores proposed actions and supplies corrective feedback while a task policy learns. The final policy runs alongside a frozen standing or walking controller, while the critic and safety controller used during training are removed.
All learning took place in MuJoCo with 4,096 parallel environments and policies running at 50 hertz. Domain randomization varied robot dynamics, actuation, contacts and outside disturbances before transfer to the physical G1. The shared failure specification covered whole-body collision, pelvis orientation and trunk height; only the object defining collision changed between incoming balls and an overhead bar.
In the first hardware dodgeball test, the real robot received its measured body state together with scripted ball positions. LIMBO's trained policy and an analytical-barrier baseline each faced the same 59 virtual trajectories, making the obstacle motion repeatable. Both were hit 10.2% of the time, but LIMBO held median position drift to 0.162 metres instead of 0.944 metres and peak body tilt to 14.9 degrees instead of 32.4 degrees. With human-thrown balls tracked by motion capture, LIMBO avoided 20 of 26 throws, compared with 18 of 26 for the baseline. That hit-rate difference was not statistically significant, while the reported reductions in drift and tilt were.
The low-obstacle experiment also shows that the safety objective did not prescribe one motion. By changing how strongly training sampled the estimated boundary of recoverability, the same specification produced strategies ranging from crouching to a backward-leaning limbo motion. The paper reports hardware deployment without an online filter, but its detailed clearance-rate analysis is primarily simulation evidence and the physical result is presented through demonstrations.
LIMBO therefore demonstrates a route for moving a learned safety structure out of the deployment loop and into policy training, reducing runtime machinery on the tested robot. It does not certify arbitrary humanoid behavior: the safety definition was chosen by the researchers, training depended on simulation and domain randomization, the hardware trials were small, and the results have not been independently replicated.
Sources
- LIMBO: Learning and Internalizing Model-Free Barrier Objectives for Agile and Safe Whole-Body Control, arXiv:2609.22075v1, September 18, 2026.
- Full paper in HTML, including training conditions, simulation results and hardware tests.
- LIMBO project page and demonstrations, research team.
- Lead image: Unitree robots at the Japan Mobility Show 2025 by RuinDig/Yuki Uchida, licensed under CC BY 4.0; unmodified. The photograph is illustrative and does not show the reported experiment.