Deploys trained Reinforcement Learning (RL) policies to real robots using high-performance Rust and ONNX Runtime.
Kinfer Runtime is a specialized model inference engine designed for the K-Scale robotics ecosystem, enabling the deployment of neural network policies directly onto physical hardware. By leveraging Rust and the ONNX Runtime (ort), it provides a low-latency pipeline essential for real-time robotics applications where timing is critical. It features optimized model loading, async telemetry logging via NDJSON, and seamless integration with KOS firmware, making it an essential tool for developers bridging the gap between simulated RL training and real-world robot execution.
Key Features
012 GitHub stars
02Cross-language support with Python bindings provided via PyO3.
03High-performance ONNX model inference using Rust and the ort-sys backend.
04Efficient asynchronous execution using a dedicated Tokio runtime with GIL management.
05Asynchronous NDJSON telemetry logging for non-blocking performance debugging.
06Real-time sensor-to-actuator execution loop optimized for embedded systems.
Use Cases
01Executing real-time neural network inference on resource-constrained embedded hardware.
02Deploying walking or manipulation policies from PyTorch/ONNX to physical humanoid robots.
03Collecting high-frequency telemetry and sensor data for policy performance analysis.
