Sub-Millisecond Obstacle Avoidance
Harware Acceleration and Vision Pipeline optimization. Klyff Engineers know hardware specific SDKs to solve the most compelling problems for the best ROI
The Problem
A warehouse robotics company was using NVIDIA Jetson Orin modules but couldn’t achieve the 60 FPS required for high-speed navigation while running three different neural networks (depth, object detection, and path planning).
Solution
The “Hard Tech” Solution: Klyff engineers performed layer-wise INT8 quantization and wrote custom TensorRT plugins to optimize the memory pipeline. We moved the pre-processing of image data from the CPU to the GPU/NPU using specialized GStreamer pipelines.
Hardware
- Main SoC: NVIDIA Jetson AGX Orin 64GB (the gold standard for high-performance edge AI).
- Co-Processor: Hailo-8 AI Accelerator (delivering 26 TOPS at only 2.5W).
- Vision: Stereo Depth Cameras (e.g., Luxonis OAK-D) or Ouster LiDAR sensors.
- Â
Software
Optimization: TensorRT (NVIDIA’s inference engine) and Hailo Dataflow Compiler.
Pipeline: NVIDIA DeepStream SDK (for GStreamer-based multi-stream processing).
Middleware: ROS 2 (Robot Operating System) for the control loop, optimized with DDS (Data Distribution Service) for low-latency communication.
Results
- Drop in latency from 45ms to 12ms
- Robots move 3x faster through the warehouse safely
- Minimal new hardware required for maximum jump in efficiency
- 24/7 support provided by Klyff engineers to ensure smooth functioning
Solution Flow
Other Case Studies
Selected Customer Success Stories
Adaptive Predictive Maintenance
Solder Joint Inspection
