Pigey bridges the orchestration gap, boosting robot reasoning 4x without retraining

A new paper from Stanford researchers introduces Pigey (Physical Agency orchestrator), a system that addresses what the authors call the 'orchestration gap' in generalist robots. Rather than training monolithic policies to reason end-to-end via massive pre-training, Pigey decomposes robotic intelligence into two separate layers: a high-level agent manager that handles planning, subgoal decomposition, success verification, and failure recovery, and a low-level language-conditioned policy or skill set that executes motor commands. This closed-loop architecture allows Pigey to orchestrate existing vision-language-action (VLA) policies without any additional data collection or post-training.

The results are striking. On the LIBERO-PRO benchmark, enabling the frozen policy with Pigey's orchestration improves success rates from 12.8% to 53.3%—a more than 4x improvement. On a real robot platform, the system lifts performance from near-zero to over 90% on reasoning-limited tasks. Pigey can switch between multiple low-level controllers (both learned VLA policies and hand-engineered parametrized skills), checks observations against planned outcomes, and recovers autonomously when actions fail. The work demonstrates that better orchestration—not larger training datasets—may be the key to unlocking general-purpose robot reasoning.