What NVIDIA Ising Taught Us About the Future of Physical AI

On April 14, 2026, NVIDIA publicly released Ising — the world's first family of open-source AI models purpose-built for quantum processor development. Ising has two components: Ising Calibration, a vision-language model that automates quantum processor calibration (cutting time from days to hours), and Ising Decoding, 3D convolutional neural networks for quantum error correction (up to 2.5x faster, 3x more accurate than the pyMatching baseline). NVIDIA explicitly framed Ising as "the control plane for quantum machines" — an AI-based validation and calibration layer that makes noisy quantum processors practical. The models are fully open-source on GitHub and Hugging Face, with 15+ organizations already adopting them.

We have been running our own validation experiments on robot foundation models. Here is what we found: NVIDIA's GR00T N1.6 (3B parameters), fine-tuned on the official LIBERO-Spatial dataset for 20K steps on an H100 SXM 80 GB GPU, achieves 97.65% success rate on LIBERO (MuJoCo). This is NVIDIA's official benchmark result for this model class. The same checkpoint, evaluated on RoboGate's 68 industrial Isaac Sim scenarios, scores 0/68 (0% success rate, Confidence Score 49/100). All 68 failures are grasp misses — zero collisions, zero drops, zero timeouts. This 97.65 percentage point gap is not a fine-tuning failure. Our scripted IK controller scores 100% on the identical 68 scenarios, confirming the evaluation pipeline works correctly. The gap arises from systematic differences between simulators: rendering pipeline, physics solver, object meshes, and lighting model.

Quantum processors are noisy — qubits decohere, error rates hover around 10^-3, and every processor requires continuous recalibration. Ising exists because without a validation and calibration layer, quantum hardware is impractical. Physical AI has the same structural problem. Robot foundation models are trained on specific simulators (MuJoCo, PyBullet) with specific rendering, physics, and object distributions. When these models encounter a different environment — even a different simulator, let alone the real world — performance can collapse completely. The 97.65% to 0% gap we measured is the Physical AI equivalent of quantum decoherence: the model works perfectly in its native environment and fails completely outside it.

We are preparing the following updates: • Paper v3: 9-model VLA leaderboard, cross-simulator gap analysis, 4-robot cross-embodiment study (Franka Panda + UR3e + UR5e + UR10e) • Community benchmark: Contributing RoboGate's 68 scenarios to Isaac Lab-Arena • More models: Cosmos Policy (NVIDIA, 2B) and GR00T N1.7 evaluations in progress If you are developing or deploying a VLA model, we encourage you to evaluate it on RoboGate's 68-scenario suite before making deployment decisions based on single-benchmark results.