PGX framework pairs humanoid VLA models with classical control layers for long-horizon tasks
Researchers behind the PGX framework say splitting temporal and precision control away from a vision-language-action model improves stage progression and success rates on long-horizon humanoid tasks.
In plain words
Vision-language-action models let robots decide how to act, but the paper's authors say these models struggle when a task requires many steps in a precise order. PGX is a framework that hands timing and precise positioning to simpler, non-learning components, so the learned model only handles the manipulation skills within each step. It was tested on a Unitree G1 humanoid in simulation and one real-world task.
What was announced
PGX, described in a paper by researchers at KL Research, organizes a long-horizon loco-manipulation task into three layers. A predicate-based state machine uses physical conditions to decide when phases begin, trigger recovery, and coordinate navigation and waiting. Vision-language-action skills then run during active manipulation phases, with flat or stage-trained models focusing on task-relevant actions. A third layer, world-coordinate locking, maintains spatial consistency through base station keeping, object-relative positioning, and world-anchored contact targets.
The authors report experiments on three simulated tasks and one real-world task with a Unitree G1 humanoid, stating that PGX improved stage progression and task success rates. The abstract does not include numerical results, and the findings are self-reported rather than independently verified. A project page accompanies the paper.
What it does not show
- The evidence is limited to the paper's abstract, so no numerical results, baselines, or failure cases can be assessed from the supplied passages.
- The reported improvements come from the authors' own experiments and have not been independently confirmed.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.