GAE paper details generator–executor learning and anticipation trade-offs
ArXiv v1 describes each approximately852M policy, PPO generator–executor training, simulation tracking results and separately scoped real demonstrations. Anticipation0–5steps at50Hz trades tracking precision against0–100ms configured lead time; it is not measured end-to-end latency.
Previously reported
The author project already documented the2025 public teleoperation demonstration, without the current paper’s exact training and quantitative evaluation record.
This announcement
ArXiv v1 describes each approximately852M policy, PPO generator–executor training, simulation tracking results and separately scoped real demonstrations. Anticipation0–5steps at50Hz trades tracking precision against0–100ms configured lead time; it is not measured end-to-end latency.
Ordinary: technical documentation of an existing model, without verified new public weights, training-code access or a major independently established capability release. Some errors worsen against SONIC, and model/runtime identity with licensed commercial software is unconfirmed. Admission score95 does not determine this editorial decision.
