Robots deserve their own inference stack.
Seven principles behind the NiLabs engine.
-
The model belongs on the robot.
A robot that has to ask a datacenter what to do next is a robot on a leash. Intelligence that matters in the physical world has to run where the physical world is: on the chip inside the machine, within its power budget, without a network.
-
Robots are not chatbots.
Chatbot serving is built for throughput: big batches, variable lengths, averages. A robot runs one request at a time, of fixed shape, against a clock. If a model emits k actions per call and the robot runs at f Hz, every call has to finish within k / f. At one action and 30 Hz, that is 33 milliseconds. Tools built for the other workload do not solve this one.
-
The worst case is the only case.
An average latency is a statistic. A slow call is a collision. We measure, optimize and guarantee the tail, because that is where the robot lives.
-
Faster is worthless if the robot stops folding the shirt.
Every quantized weight, pruned layer and kernel trick can quietly change what the robot does. Speed claims without task success are marketing. We report both, every time.
-
Any chip.
The best model should not be locked to one vendor’s silicon. Robot builders should pick a chip for cost, power and supply, and expect today’s models to run well on it.
-
Open by default.
The engine is open source under Apache-2.0. Benchmarks are reproducible, with the code and the method published. If you cannot rerun our numbers, you should not trust them.
-
New models, running in days.
Open VLAs improve every month. Our job is to make each new release run on real robots within days, not after a quarter of hand-tuning.
Physical AI is leaving the lab. We are building the layer that lets it run on the robot.