arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

PGX boosts success from 35% to 80% on coffee‑fetching for GR00T and from 0% to 60% for π₀.₅ | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
RoboticsEnglishPublished

PGX boosts success from 35% to 80% on coffee‑fetching for GR00T and from 0% to 60% for π₀.₅

November 22, 2026arXiv: 2611.26596v1

Researchers present a physics‑based execution system called PGX that splits long‑horizon tasks into a predicate‑based state machine, a vision‑language‑action (VLA) skill layer, and a world‑coordinate locking (WCL) module. This design lets the VLA focus on the right actions for each stage while classical elements handle timing and precision.

On a coffee‑fetching job, the PGX framework lifted success from 35 % to 80 % for the GR00T policy and from 0 % to 60 % for the π ₀.₅ policy without changing neural weights. Similar gains appeared in all three simulated scenarios—box transport, mug placement, and coffee fetching—where average success rose from 35 % to 80 % with GR00T and from 17 % to 82 % with π ₀.₅.

The improvements stem from explicit gating of stages by physical predicates. The system can start the next phase only after a predicate such as “mug lifted” evaluates true, eliminating the ambiguity that baffles end‑to‑end models.

However, PGX introduces extra time for stage transitions because the predicate checks must run before each skill. Moreover, the approach relies on accurate state sensing—errors from cameras or limb encoders can trigger wrong transitions. Finally, the framework requires designers to define task‑level predicates and stage boundaries, which may limit its applicability to truly new tasks.

PGX shows that separating control into classical, physics‑grounded layers can substantially improve robot performance, but practical deployment will need reliable state estimates and careful task definition.

Source paper

PGX: Physically Grounded Execution for Long-Horizon Humanoid Loco-Manipulation

RoboticsArtificial Intelligence
View on arXiv