👋 About Me
Hi everyone! I’m David, and I’m excited to take part in another Axelera AI challenge.
Some of you may remember me from Axelera’s previous Smarter Spaces competition, where my project was selected as the winning project. That experience gave me the opportunity to explore what Metis can do for real-world edge AI.
For The Prompt, I want to take Metis in a completely different direction: from a smart-home system to the perception engine of an autonomous vehicle.
🏭 The Challenge
Autonomous vehicles such as AGVs, warehouse robots and industrial rovers need to continuously understand their surroundings.
A single camera provides only a limited view. Adding more cameras gives the vehicle significantly more visual information, but it also increases the amount of video that needs to be processed in real time.
This creates an interesting edge-AI challenge:
Can a single Metis accelerator process multiple live camera streams and turn them into useful perception for a moving vehicle — all locally and with low enough latency for real-time control?
That is what I want to find out.
🚗 The Project — SafeTrack Rover
I will build a small autonomous rover equipped with** Wi-Fi cameras positioned at different angles**, providing complementary views of the environment.
The rover will continuously stream its camera feeds over the local network to a Dell Pro Slim Plus XE5 equipped with a 16GB Axelera Metis PCIe accelerator.
The XE5 will act as the rover's external edge-AI brain.
The system will combine two vision tasks:
🎯 Target Detection & Tracking
A YOLO-based model will detect and track a selected target.
The rover will use the target's position to determine the direction it should move and attempt to follow it autonomously.
👁️ Spatial Awareness
A depth-estimation model will process the camera streams to estimate the distance of objects in the rover's environment without expensive LIDAR sensors.
Instead of simply displaying a depth map, the system will convert the depth information into a compact spatial representation around the rover.
For example:
CLEAR → CAUTION → STOP
The perception system will continuously combine target information and obstacle information to generate steering and braking commands.
These commands will then be sent back to the rover through a local HTTP API.
The result is a complete perception-to-action loop:
Cameras → Metis → AI perception → spatial reasoning → steering → rover
🪽 The Voyager Wingman Challenge
The most important part of this project is not simply building an RC rover.
I want to use Voyager Wingman as the AI pipeline engineer.
I will start with a natural-language description of the system rather than manually designing the complete pipeline myself.
My initial prompt will be:
"Act as an edge-AI robotics architect. I am building a small autonomous rover equipped with Wi-Fi IP cameras positioned at different angles. The cameras provide continuous RTSP or HTTP video streams to an Axelera Metis system. Design a real-time multi-camera perception pipeline using the Voyager SDK. The pipeline should concurrently process both camera streams and perform two tasks: target detection and tracking using a YOLO-family object detection model, and monocular depth estimation for spatial awareness. Keep inference on the Metis accelerator wherever supported. Generate the pipeline configuration and Python application logic required to ingest the streams, associate model outputs with their camera sources, extract useful depth information around the rover, and produce real-time steering and braking decisions. The resulting commands should be represented as JSON and sent to the rover through a local HTTP API. Prioritize low latency, concurrent processing, robustness to multiple video streams, and measurable performance. Start with the simplest working pipeline and iteratively optimize it."
From there, I will let Wingman generate, debug and optimize the pipeline.
⚙️ The One-Month Build
Week 1 — Make It See
-
Set up the Dell XE5 and Metis environment
-
Connect the first IP camera
-
Generate the initial Wingman pipeline
-
Deploy object detection
-
Deploy depth estimation
-
Establish communication between the XE5 and rover
Week 2 — Add the Second View
-
Introduce the second camera
-
Run the streams concurrently
-
Associate inference results with their camera sources
-
Develop the spatial-awareness logic
-
Measure FPS, latency and resource utilization
-
Use Wingman to troubleshoot and optimize the pipeline
Week 3 — Make It Move
-
Integrate the motor controller
-
Implement target following
-
Implement obstacle-aware steering
-
Add emergency braking
-
Test the rover in progressively more challenging environments
Week 4 — Push It Further
-
Optimize the final pipeline
-
Benchmark camera configurations
-
If the system is stable, add a third camera as a stretch goal
-
Measure how the additional stream affects throughput and latency
-
Prepare the final demonstration and documentation
📊 The Experiment
I don't just want to demonstrate that the rover works.
I want to measure how the Metis system scales as the visual workload increases.
The core experiment will compare:
1 camera → 2 cameras
with a possible:
3-camera stretch configuration
For each configuration I will measure:
-
FPS per camera
-
End-to-end perception latency
-
Detection latency
-
Depth-estimation performance
-
CPU utilization
-
Metis utilization
-
Target-following performance
-
Obstacle-response latency
This should provide a practical demonstration of the capability of Metis for multi-camera edge AI workloads.
🛡️ The Final Demo
The final demonstration will show the rover receiving multiple live camera streams over Wi-Fi while the Dell XE5 processes the video locally.
The final video will show both sides of the system simultaneously:
the rover physically navigating the environment + the live AI perception running on the Metis system.
🚀 Why SafeTrack Rover?
For me, this project is an opportunity to explore a completely different application of Axelera AI.
My previous project explored multi-camera AI for interaction with a smart environment.
Now I want to put that perception capability on a moving machine, where minimal latency is critical.
The rover itself is intentionally simple.
The interesting part is what happens inside the edge-AI system:
multiple live video streams, multiple AI models, real-time spatial reasoning and a physical control loop — generated and iteratively optimized through natural-language interaction with Voyager Wingman.
The core question is simple:
How far can one Metis accelerator take the perception of a small autonomous vehicle?
And with Voyager Wingman as the development wingman, I want to find out in just one month.
I'm ready to build. 🚀
