Free for humans

Long humanoid jobs, split between a skill model and physical rules

Physically Grounded Execution leaves the current skill to a vision-language-action model and moves timing and precision onto two classical training-free components. On three simulated tasks and one Unitree G1 task, stage progression and success improve, with no rates given.

arXiv:2611.265965 min readScore 76/100 · editorial triage · not peer reviewPaper hub2026-W48

The 30-second take

  • Vision-language-action models can run individual manipulation skills, but long loco-manipulation still couples timing, skill, and precision inside one model that struggles.
  • A predicate state machine gates phases, recovery, navigation, and waiting, while world-coordinate locking holds the base, the object, and the contact target consistent.
  • Experiments on three simulated tasks and one real Unitree G1 task show better stage progression and task success, and the abstract does not quantify either gain.

What the paper actually did

Physically Grounded Execution, or PGX, is a framework for long-horizon humanoid loco-manipulation. It offloads temporal control and precision control to two classical training-free components so a vision-language-action model can focus on stage-specific skill dynamics. The abstract decomposes the system into three layers. A predicate-based state machine uses physical predicates to gate phase transitions, trigger recovery, and coordinate navigation and waiting. The same predicate semantics are shared by offline segmentation of demonstrations, online execution, and stage-wise evaluation. Vision-language-action skills then operate in the active manipulation phases, using either a flat model or a stage-trained model concentrated on the relevant actions. World-coordinate locking maintains spatial consistency through base station keeping, object-relative positioning, and world-anchored contact targets. Experiments cover three simulated tasks and one real-world task on a Unitree G1 humanoid. The abstract reports improvements in stage progression and task success rates.

What makes this disruptive

The usual pressure on long robot jobs is to absorb horizon, skill, and precision in one larger model. PGX does the opposite. It keeps the learned model inside the phase where a skill is happening, and it assigns timing and precision to classical components that do not require training. The predicate layer is the widest conceptual change. One set of physical predicate meanings is asked to cut the demonstrations, steer the live robot, and score the stages. World-coordinate locking then keeps the base, the object-relative pose, and the contact target from drifting into inconsistent frames. The abstract claims this split improves stage progression and task success. It does not claim that the learned model became unnecessary.

Why it matters (outside the lab)

A humanoid that only succeeds at isolated hand skills cannot finish a job that requires going somewhere, waiting, recovering, and making contact in a stable place. Loco-manipulation is that longer chain, and it is the kind of work people mean when they talk about a robot leaving a bench and completing a task. Because phase changes are gated by physical predicates, a reviewer can ask which predicate allowed the next phase, which is a different question from asking what action the network emitted. The evidence in the abstract is three simulated tasks and one real task on a Unitree G1. That is enough for the authors to report improvement, and the project page is public for inspection.

Limitations & open questions

The abstract does not report a success rate, a stage-progression rate, or the size of either improvement. It does not name the three simulated tasks or the real-world task. It does not list the predicates, say how they were chosen, or say what happens when a predicate is wrong. It does not compare flat models with stage-trained models using any number. It does not present a robot other than the Unitree G1. It does not show that the training-free components transfer to a new job without new predicate design. Improvement is a real claim in the abstract, and it is not a numeric claim. The abstract also does not further define base station keeping.

Explain ladder

Default article depth

Models that look at a scene and output a robot action are good at single manipulation skills and weak when walking, waiting, recovery, and precise contact all have to stay in step. This framework gives the timing and the recovery rules to a state machine based on physical yes-or-no checks, gives precision to a spatial lock, and lets the learned model concentrate on the skill of the current phase. Tests include a real Unitree G1 humanoid, but the abstract does not say by how much success improved.

Key terms

Loco-manipulation
A task that couples locomotion with manipulation over a long horizon. The abstract says monolithic vision-language-action models still struggle when timing, skill, and precision are bound together in that setting.
Physical predicate
A yes-or-no statement about the physical situation, used here to gate phase changes, trigger recovery, and coordinate navigation and waiting. The abstract does not list the predicates.
World-coordinate locking
The paper's mechanism for spatial consistency, using base station keeping, object-relative positioning, and world-anchored contact targets. The abstract does not further define base station keeping.
Vision-language-action model
A learned model that maps visual and language context to robot actions. In this framework it is used for stage-specific skills, either as a flat model or as a stage-trained model, while timing and precision sit elsewhere.

Sources

Related explainers

Same topic and week first — keep exploring the scarcity → abundance map.

Editorial explainer · not peer review · always read the primary paper.

Byline: Disruptive Concepts editorial.