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

Kai Lu1 Chenyang Ma2 Ioannis Havoutis1

1Oxford Robotics Institute, University of Oxford 2Department of Computer Science, University of Oxford

A structured execution framework combining predicate-based task progression, VLA skills, and world-frame spatial stabilization.

Abstract

Vision-language-action (VLA) models can execute individual manipulation skills, but long-horizon humanoid loco-manipulation also requires reliable task progression and spatial precision. Physically Grounded Execution (PGX) separates these responsibilities into three layers: a predicate-based state machine controls transitions, recovery, navigation and waiting; the VLA executes skills within active manipulation stages; and world-coordinate locking (WCL) maintains spatial consistency through station keeping, object-relative positioning and world-anchored contact targets.

Method

PGX decouples timing and world-frame precision from the learned policy, allowing the VLA to focus on transferable manipulation skills. Physical structure determines when and where to act; the VLA learns how.

PGX architecture: predicate-gated state machine, VLA manipulation skills, and world-coordinate locking, evaluated on four tasks.
Framework overview. A predicate-gated state machine governs progression; the VLA executes manipulation skills; WCL supplies spatial stabilization.

01 Timing

Predicate-gated state machine

Enter the next stage when its physical guard is satisfied. Timeouts trigger retries, while reverse transitions recover from premature progression. Navigation and waiting are handled by the executor.

02 Skills

Vision-language-action policy

Execute manipulation skills from an executor-prepared start state. PGX supports flat or stage-trained VLA skills; stage-wise training is optional and its effect is model-dependent.

03 Precision

World-coordinate locking

Compute navigation goals from the current object pose, hold the base at its manipulation-entry position, and anchor precise contact targets in the world frame.

Experimental results

Experiments use a Unitree G1 in Isaac Lab Arena, with at least 20 evaluation episodes per cell. We report end-to-end task success separately from intermediate stage progress.

Execution structure with fixed policy weights

The first two result columns below keep the VLA weights and task-level instruction fixed. The final column uses a stage-trained skill model. This comparison isolates the effect of the execution framework from that of skill training.

Task success (%) (Table III)
Backbone / conditionFlat executionPGX + flat skillsPGX + stage skills
π₀.₅ (C0)06060
π₀.₅ (C1)06055
GR00T N1.6 (C0)358035
GR00T N1.6 (C1)208035

C0: nominal initial pose. C1: randomized initial pose. Stage-wise training is model-dependent; flat GR00T skills perform better on coffee fetching in this evaluation.

Stage progression on coffee fetching

Stage survival for pi 0.5 and GR00T N1.6 comparing Flat VLA, FSM only, PGX, and PGX with WCL-P.

Flat-policy failure examples

π₀.₅
GR00T N1.6

Percentage of episodes reaching each stage with π₀.₅ (top) and GR00T N1.6 (bottom).

PGX success examples

Coffee fetching (π₀.₅)

PGX executes the coffee task with navigation, machine interaction and delivery.

Coffee fetching (GR00T N1.6)

A second VLA backbone executes the task within the PGX framework.

Task success across three tasks

Task success (%) (Table II)
TaskBackboneNominal (C0)Random initial pose (C1)
FlatPGXFlatPGX
Box transportGR00T N1.655752045
π₀.₅1090560
Mug placementGR00T N1.61565570
π₀.₅075090
Coffee fetchingGR00T N1.635802080
π₀.₅060055

C1 adds uniform offsets of ±0.30 m in both translation axes and ±30° yaw. Time budgets are 40 s, 60 s and 110 s for box, mug and coffee, respectively. PGX uses stage-trained skills for box and mug; coffee uses stage-trained π₀.₅ and flat-trained GR00T. These comparisons do not all hold policy weights fixed.

World-frame stabilization

WCL reduces press-stage p90 maximum base drift from 54 to 9 cm and increases completed placement attempts from 74 to 97 percent.
With WCL, completed placements rise from 74% to 97%, and press-stage p90 maximum base drift falls from 54 cm to 9 cm.
Base trajectory under injected disturbance. With and without WCL under sustained disturbance (3× speed).

Real-robot experiments

PGX is deployed on a Unitree G1 with Dex-3 hands. The VLA is fine-tuned on 44 real demonstrations; motion capture and joint states supply physical predicates. Simulation and real-world policies are trained separately.

Hardware setup

Unitree G1 with Dex-3 hands, head and wrist cameras, mug, coffee machine and delivery area.
Unitree G1 with Dex-3 hands and head and wrist cameras.

Stage-wise completion

Per-stage completion in simulation and on the real G1, for grasp, place, re-grasp and delivery.
Simulation and hardware completion rates, conditioned on reaching each stage. The real-robot task does not include button pressing.
Real Unitree G1 coffee task. PGX with the π₀.₅ stage model, shown at 4× speed. The stage bar identifies navigation, manipulation and waiting. A 2 s freeze-frame highlights the wait stage.

Task demonstrations

PGX executes long-horizon coffee fetching, box transport and mug placement in simulation.

Coffee fetching (π₀.₅)

Head-camera view of PGX execution (4× speed).

Coffee fetching (GR00T N1.6)

Head-camera view of PGX execution (4× speed).

Box transport

Bimanual carrying and placement in the target area.

Mug placement

Grasp, transport and placement of the mug in the target tray.

BibTeX

@article{lu2026pgx,
  title   = {PGX: Physically Grounded Execution for Long-Horizon Humanoid Loco-Manipulation},
  author  = {Lu, Kai and Ma, Chenyang and Havoutis, Ioannis},
  journal = {arXiv preprint arXiv:2611.26596},
  year    = {2026}
}