Run robot foundation models on any robot chip.
NiLabs builds an open-source inference engine for vision-language-action (VLA) models. Faster, lower power, real-time, and verified to keep task success.
Robotics is moving to large learned models such as OpenVLA, SmolVLA, π0 and GR00T. They are too slow and too power-hungry for the chips inside robots, so teams hand-optimize for weeks, tether robots to big GPUs, or ship jerky behavior.
Robot inference is not chatbot serving. It runs at batch size 1, on fixed-length action chunks, against a hard clock: a model that emits k actions per call on a robot running at f Hz has k / f to produce the next chunk. At one action and 30 Hz, that is 33 ms, on every call.
- Optimized code for each stage of a VLA: vision encoder, language backbone and action head
- Asynchronous inference, so the robot keeps moving while the next action chunk is computed
- Quantized model formats and a one-command converter from Hugging Face and LeRobot checkpoints
- Published benchmarks for latency, control rate and power on every supported chip, reported next to task success
Pre-release. The first target is one open VLA on one chip, benchmarked against stock PyTorch with task success matched. Code, benchmarks and the method will be published here under Apache-2.0.
We're looking for design partners: robot teams running VLAs onboard, and chip makers who want VLAs to run well on their silicon. Email hello@nilabs.dev.
