FARO asks an LLM for a humanoid's contact plan, then checks if the motion is physically possible

Picking up a box and putting it on a shelf sounds like one task. For a humanoid, it is two at once: the robot has to walk somewhere with the box, then place it, while staying balanced on legs that have fewer actuators than degrees of freedom. Existing planners make that trade-off by simplifying the robot model or handing engineers a library of pre-baked motion primitives. Reinforcement-learning controllers, the alternative, need demonstrations to imitate and cannot invent new behaviors when the scene changes.

A new arXiv preprint from TUM and CMU, FARO: Feasibility-Aware Robot Motion Optimization, tries a different split. The framework breaks humanoid loco-manipulation into three stages, and the paper's HTML version walks through each one.

First, a Large Language Model proposes a sequence of contacts the robot should make with the world: a hand here, a foot there, then a grasp. The LLM is the contact suggester, not the planner. Second, a feasibility-guided tree search takes those suggestions and throws out the ones that do not admit a physically executable motion given the robot's dynamics. The search is the gate. Third, a reinforcement-learning controller tracks the motion that survives the gate, in simulation and on hardware.

That second step is the contribution the paper leans on. Tree searches in robotics are not new, but pairing one with an LLM-sampled contact space is the structural novelty. The authors frame their pipeline as filling an exploration gap that both classic trajectory optimization and demonstration-driven RL leave open.

The concrete claim, per the paper and a supplementary video, is that generated trajectories are "of sufficient quality for execution in real-world loco-manipulation scenarios." That wording is the authors', and it is the right one to flag: the video is author-supplied, there is no independent benchmark or replication in the public record, and the preprint has not been peer reviewed. The work is also a single paper, posted to arXiv on 2026-07-20, with no deployment, partner, or product claim.

The authors are Michal Ciebielski, Shafeef Omar, and Majid Khadiv at TUM MIRMI, plus Aaron Johnson at TUM's Institute for Advanced Study and Carnegie Mellon Mechanical Engineering. Funding comes from Siemens AG, the TUM Institute for Advanced Study, and the Huawei-TUM Joint Laboratory, which is worth knowing because industrial-academic sponsorship can shape benchmark choices.

A fair counter is that a tree-search gate is only as useful as the contact sequences the LLM proposes in the first place. If the model keeps suggesting the same five contacts, the search will keep rejecting the same five. The real test, which the preprint does not run, is how the contact space grows as the tasks get messier and how the pipeline compares to a major lab's humanoid planner on a shared benchmark. Watch item: a peer-reviewed venue acceptance, or a third-party reproduction.