ROBBRIEF
A robotics preprint proposes splitting a robot's reflexes from its planning with a switch that picks which AI to call, and reports 93 control decisions per second on a standard simulated test.
A new robotics preprint says a two-armed robot can make about 93 control decisions per second, roughly 50% faster than a typical 60-frame-per-second video game, not by running a bigger AI, but by running two AIs and switching between them on the fly.
The paper, "Fast and Accurate: An Adaptive VLA Inference Framework through Environment-aware Model Selection," describes a design the authors call Environment-aware Model Selection. Instead of one robot AI handling both fast reactions and slow planning, the system pairs a lightweight controller for routine moves with a larger, more deliberate model that only kicks in when something unfamiliar shows up. A reinforcement-learning switch watches the situation in real time and decides which model to invoke, replacing the tightly coupled hierarchical pattern that most robot stacks use today.
On LIBERO, a standard simulated benchmark for lifelong robot learning, the authors report their design matches the success rate of a large-scale baseline while running at 93.4 Hz. They also describe real-world two-armed tasks where the small model handled most moves and the big planner stepped in for harder ones, with the authors reporting faster completion and maintained robustness.
The source is an arXiv preprint, not peer-reviewed. The 93.4 Hz figure is compared only against a single large baseline; there is no independent reproduction, and the real-world claim is qualitative from the authors' own lab. The next test will be an independent LIBERO reproduction and a per-task breakdown showing how often the switch wakes the slow model.