NVIDIA's Metatron v1 is downloadable. That does not prove full-shift robot autonomy.

Open weights are particularly meaningful for robotics because they let outside engineers inspect model inputs, outputs and program interfaces rather than judging capabilities from video. NVIDIA's NVlabs ProtoMotions project announced Metatron v1 on 28 October, describing a learned Unitree G1 policy spanning locomotion, steering, scene interaction and recovery after a fall.

The Hugging Face model card distinguishes three downloadable artifacts: a roughly 459 MB native PyTorch checkpoint, a roughly 403 MB ONNX browser export for one robot, and a similarly sized ONNX network for a batched six-robot simulator. They share learned weights but not interchangeable input contracts. Native checkpoint information and browser deterministic inference paths serve different use cases.

This is stronger transparency than promising weights in a future release. The repository publishes byte counts, SHA-256 hashes and interface documentation, which allow reproducible download and version pinning. Yet a matching file checksum proves the artifact's identity, not independent evidence of how well a robot performs in an unfamiliar factory.

Motion control is difficult because humanoids must balance, steer, respond to changing goals and recover from perturbations. A unified policy could reduce hand-offs between specialized controllers. Metatron's model card documents 29 G1 controlled joints, observation histories, task goals and control outputs that real applications must integrate correctly.

The maintainers describe action-parity testing between browser exports on 60 distinct and perturbed simulator states, with a reported maximum action difference of 5.55e-6. This is a useful software consistency measurement. It is not a success rate of 60 robot tasks, a hardware safety certification or evidence of eight hours of independent work.

Licensing requires equally careful wording. ProtoMotions source code is Apache-2.0; the model files follow NVIDIA's Open Model License. Robot assets, motion clips and third-party datasets keep separate rights. Researchers can download the weights, but a fully reproducible training corpus is not included in this model repository.

Cortist's F-003 forecast requires repeated, independently checked eight-hour industrial shifts across mixed tasks without hidden teleoperation by 2035. Metatron's release does not attempt that test. It supplies potentially useful control infrastructure, not a resolved outcome or reason to adjust the provisional forecast probability.

The decisive experiment would use a physical G1 under external supervision, reporting completed tasks, resets, interventions, falls, collision risks, unexpected environments, uptime and recovery behavior. Simulation demos and numeric output parity are informative engineering prerequisites; they remain separate from real industrial shift evidence.