PGX: Physically Grounded Execution for Long-Horizon Humanoid Loco-Manipulation
Organizations: K. Lu and I. Havoutis are with the Oxford Robotics Institute, University of Oxford. · C. Ma is with the Department of Computer Science, University of Oxford.
Abstract
Robotic vision-language-action (VLA) models can execute diverse individual manipulation skills, yet they still struggle with long-horizon loco-manipulation tasks, where temporal structure, skills, and precision are tightly coupled within a monolithic model. We propose Physically Grounded Execution (PGX), a framework that offloads temporal and precision control to two classical training-free components, allowing the VLA to focus on stage-specific skill dynamics. PGX decomposes long-horizon loco-manipulation into three layers: (i) a predicate-based state machine governs task progression, using physical predicates to gate phase transitions, trigger recovery, and coordinate navigation and waiting. Shared predicate semantics connect offline demonstration segmentation, online execution, and stage-wise evaluation; (ii) VLA skills operate within active manipulation phases, with flat or stage-trained models concentrating on task-relevant actions; and (iii) world-coordinate locking (WCL) maintains spatial consistency through base station keeping, object-relative positioning, and world-anchored contact targets. We demonstrate the effectiveness of PGX with extensive experiments conducted on three simulated tasks and one real-world task using a Unitree G1 humanoid robot for long-horizon loco-manipulation, showing improvements in stage progression and task success rates. Project page is available at: https://kl-research.github.io/pgx
Figures & tables
| Nominal (C0) | Random init. (C1) | ||||||
| Model | Lift | Press | Succ. | Lift | Press | Succ. | Avg. |
| Flat VLA | 20 | 0 | 0 | 10 | 0 | 0 | 5.0 |
| FSM only | 92 | 48 | 32 | 95 | 55 | 50 | 62.0 |
| PGX | 100 | 70 | 60 | 100 | 65 | 55 | 75.0 |
| PGX + WCL-P | 100 | 98 | 70 | 100 | 95 | 70 | 88.8 |
| Flat VLA GR00T | 70 | 35 | 35 | 40 | 25 | 20 | 37.5 |
| Box transport | Mug placement | Coffee fetching | |||||||||||
| Nom. | Rand. | Nom. | Rand. | Nom. | Rand. | ||||||||
| Model | Lift | Suc | Lift | Suc | Lift | Suc | Lift | Suc | Lift | Suc | Lift | Suc | Avg |
| Flat GR | 95 | 55 | 25 | 20 | 35 | 15 | 5 | 5 | 70 | 35 | 40 | 20 | 35.0 |
| PGX GR | 100 | 75 | 65 | 45 | 80 | 65 | 75 | 70 | 100 | 80 | 100 | 80 | 77.9 |
| Flat | 95 | 10 | 60 | 5 | 0 | 0 | 5 | 0 | 20 | 0 | 10 | 0 | 17.1 |
| PGX | 100 | 90 | 75 | 60 | 90 | 75 | 90 | 90 | 100 | 60 | 100 | 55 | 82.1 |
| Backbone / cond. | Flat | PGX flat | PGX stage | Avg. |
| C0 | 0 | 60 | 60 | 40.0 |
| C1 | 0 | 60 | 55 | 38.3 |
| GR00T C0 | 35 | 80 | 35 | 50.0 |
| GR00T C1 | 20 | 80 | 35 | 45.0 |
| Avg. | 13.8 | 70.0 | 46.3 | 43.3 |
Long-Horizon Robotic Manipulation
Long-horizon robotic manipulation studies how robots plan, remember, and recover across extended sequences of physical tasks.
See all 468 papers