Project background
Outdoor robots face unpredictable obstacles — people, pets, debris, terrain changes. The client needed a perception package that detected these reliably at real-world speeds and lighting conditions.
Challenge
Running modern detection models within the power and thermal budgets of a mobile robot, and fusing multiple sensor modalities so that no single sensor failure produces a blind spot. False negatives had to be vanishingly rare.
Approach & solution
We built a multi-modal stack combining stereo depth, color vision, and radar, fused into a unified obstacle representation. Models were quantized and deployed to embedded GPUs with careful latency budgeting. Edge-case training data was collected specifically for high-risk scenarios.
Results & benefits
The system detected and responded to people, animals, and unexpected obstacles reliably across pilot conditions. Latency stayed low enough for the downstream planner to stop or reroute within safe distances.
