NVIDIA has unveiled ENPIRE, a novel research initiative led by Senior Research Scientist and Head of Embodied AI Jim Fan, designed to automate the experimental development of physical AI systems. The project deploys eight AI coding agents paired with eight robotic workstations, each equipped with dual six-degree-of-freedom mechanical arms, Intel RealSense depth cameras, and local NVIDIA RTX 5090 GPUs. Operating entirely on-premises, the system enables agents to independently design experiments, write and debug code, execute physical tasks, and iteratively optimize robot policies, with human researchers limited to setting objectives and monitoring outcomes. The ENPIRE framework operates in two distinct phases. Initial setup requires human engineers to establish safety boundaries, automated scene reset protocols, and verification mechanisms. Once the environmental interface is established, the coding agents transition to fully autonomous research. During development, NVIDIA evaluated Codex, Claude Code, and Kimi Code. While all three succeeded in simulation, real-world deployment exposed environmental non-determinism such as variable friction and sensor noise. Codex demonstrated superior performance, achieving target success rates with the shortest latency. To validate the platform, researchers tested high-precision physical tasks including Push-T block navigation, pin insertion, GPU socket installation, and zip tie cutting. The system autonomously explored reinforcement learning and behavior cloning strategies, adjusting hyperparameters and reward functions. This iterative process yielded a continuous success streak of fifty pin insertions. Notably, the team introduced an idea tree methodology to track hypothesis evolution, with minor architectural adjustments yielding double-digit efficiency gains and accelerating convergence beyond traditional human-led methods. A key finding from the ENPIRE trials is the scaling effect of physical compute. Increasing the robot fleet from one to eight units reduced R&D time for identical tasks by nearly half. Multiple agents simultaneously explore divergent algorithmic paths, sharing successful code via automated version control while discarding ineffective approaches. However, this parallel acceleration incurs superlinear token consumption, as agents must continuously ingest and synthesize peer discoveries. Additionally, textual experiment summaries proved effective for cross-task knowledge transfer, allowing new agents to rapidly adapt methodologies without weight retraining. Despite these advances, ENPIRE remains a prototype rather than a fully autonomous research facility. Human oversight is still required for infrastructure setup, safety configuration, and task-specific reset scripting. Current validations are confined to structured tabletop operations, leaving open questions regarding scalability to unstructured, complex environments. NVIDIA has confirmed that the underlying paper and system code will be open-sourced, signaling a strategic pivot toward automating the physical AI development lifecycle.