NiLabs
Run robot foundation models on any robot chip.
NiLabs builds an open-source inference engine for vision-language-action models. Faster, lower power, real-time, and verified to keep task success.
Each inference call has this long to produce the next actions. Every call.
- OpenVLA
- SmolVLA
- π0 / π0.5
- GR00T
- LeRobot checkpoints
- Jetson Orin
- Jetson Thor
- Edge NPUs
Robot models don’t fit on robot chips.
Robotics is moving to large vision-language-action models. They are too slow and too power-hungry for the hardware inside robots, and there is no standard tool for running them there.
Power draw of top robot GPUs. More than many robots, drones and vehicles can budget.
Time to produce one action at 30 Hz. Miss it and the robot stutters, or collides.
Of specialist hand-tuning per model, per chip. Or the robot gets tethered to a big GPU.
One engine, every robot chip.
An open-source runtime teams adopt for free, and the paid tools that take a model from a lab demo to a robot in production.
Inference engine
A runtime purpose-built for VLAs, with optimized code for each stage of the model.
- Vision encoder, language backbone and action head
- Async inference, so the robot keeps moving while the next chunk computes
- One-command import from Hugging Face and LeRobot
Optimizer
Upload a fine-tuned model, pick a target chip, get back a certified build.
- Automatic quantization, pruning and distillation search
- Task success checked in simulation
- A report of speed, power and success against the original
Production runtime
For fleets where one slow call is a safety incident, not a statistic.
- Worst-case latency guarantees
- Watchdogs, fallbacks and health telemetry
- Support contracts and ports to new chips
Built for robots, not chatbots.
Robot inference is a different workload from LLM serving. Most serving optimizations don’t transfer, so we start from the robot’s constraints.
| LLM serving | Robot inference | |
|---|---|---|
| Batch | Large, dynamic batches | Batch size 1 |
| Shape | Variable-length text | Fixed-length action chunks |
| Latency target | Good averages | Worst case, on every call |
| Success metric | Tokens per second | The robot still completes the task |
| Hardware | Datacenter GPUs | Whatever chip is in the robot |
Real-time first
Optimized for batch size 1, fixed shapes and tail latency, so the robot hits its control rate every cycle.
Any chip
One engine across NVIDIA and non-NVIDIA silicon, so teams choose hardware for cost, power and supply.
Verified task success
Every optimization is checked against the original model’s task success. Speed without it doesn’t count.
Robots deserve their own inference stack. Not a chatbot server with a robot bolted on.
Read the manifestoBring us your model. We’ll make it run on your robot.
We’re looking for design partners who want their VLAs running faster, on less power, without losing task success.
Investors: we’d like to hear from you too, at hello@nilabs.dev.
Robot teams
Running open VLAs onboard, or wishing you could. We’ll optimize your fine-tuned model for your chip.
Chip makers
Want popular VLAs to run well on your silicon. We’ll port and tune them, and publish the numbers.
Labs releasing open VLAs
Want your model running on real robots on day one. We’ll support new releases fast.
