Local Robotics Team Qualifies for FIRST Robotics World Championship in Houston

Why FIRST Team 2370’s IBOTS Are Stress-Testing Edge AI in Real-Time Control Loops

Rutland County’s IBOTS aren’t just building a robot for Houston next week—they’re deploying a hardened perception stack on NVIDIA Jetson Orin to close sensor-to-actuator loops under 8ms, a latency target that mirrors industrial cobot safety thresholds. With the 2026 FRC game emphasizing rapid object classification and dynamic path replanning, the team’s software lead replaced last year’s OpenCV pipeline with a TensorRT-optimized YOLOv8n model quantized to INT8, trading 2.3% mAP for deterministic inference. This isn’t STEM outreach; it’s a live-fire exercise in deploying ML at the edge where failure means bent aluminum, not just lost points.

The Tech TL;DR:

• Team 2370’s Jetson Orin subsystem achieves 7.2ms end-to-end latency from CSI camera input to motor PWM output under FRC 2026 game conditions.
• Their perception stack uses a custom ROS 2 Foxy node chain with DDS security profiles to prevent spoofed sensor data injection—a direct port of AGV safety patterns.
• Funding came via a $15k grant from the Vermont Advanced Computing Center, with all software committed to GitHub under MIT license as of March 2026.

The nut graf is simple: FRC has become an inadvertent proving ground for real-time AI safety practices that enterprise IT overlooks until after a breach. When IBOTS’ vision system misclassifies a retro-reflective tape target as a game piece, it doesn’t trigger a loss—it triggers a hard stop via their watchdog timer, a pattern identical to how cybersecurity auditors recommend isolating ML inference pipelines from control networks. This year’s innovation? A sidecar container running Falco to monitor for anomalous syscalls from the TensorRT runtime, logging any attempt to escape the Jetson’s cgroup—a technique lifted straight from NVIDIA’s own Jetson Linux security hardening guide.

Architecture Under the Hood: From CSI Socket to Motor Controller

Digging into the firmware, the team bypassed WPILib’s default CameraServer for a GStreamer pipeline that injects raw Bayer frames directly into a shared memory buffer, avoiding the 1.8ms copy overhead of NT4. The YOLOv8n model, pruned to 1.2M parameters, runs on the Jetson’s Orin GPU at 28 TOPS (measured via tegrastats), yielding 138 FPS at 320×240 input—enough to process two frames per control cycle. Critical path latency breakdown: 1.1ms sensor capture, 4.3ms NN inference, 0.8ms ROS 2 serialization, 0.6ms PID calculation, 0.4ms PWM output. This level of profiling would make any managed service provider managing Kubernetes clusters nod in recognition—it’s the same rigor applied to Istio sidecar latency tuning.

“We treat the Jetson like a safety-critical avionics box—every node has a heartbeat topic and if perception drops below 90% confidence for three cycles, we default to lidar-only dead reckoning. It’s not about winning; it’s about not breaking anything when the model hallucinates.”

Transparency matters: the perception stack isn’t vaporware. It’s built on publicly auditable GitHub commits, with CI pipelines running on GitHub Actions that execute unit tests against a recorded rosbag of last year’s Championship matches. Funding transparency? The Vermont Advanced Computing Center (VACC) provided not just the Jetson Orin dev kit but also access to their Slurm cluster for overnight hyperparameter sweeps—a detail confirmed in VACC’s Q1 2026 annual report. This isn’t a school project with borrowed hardware; it’s a federally supported edge AI testbed where the failure mode is a bent shooter mechanism, not a compromised SOC.

Cybersecurity Implications: When Your Perception Stack Is the Attack Surface

Here’s where it gets spicy for enterprise readers: IBOTS’ threat model assumes compromised camera firmware—a realistic scenario given the rise of supply-chain attacks on USB peripherals. To mitigate, they enforce signed firmware on the Arducam IMX477 modules via SHA-256 hashes burned into the Jetson’s eFUSE during imaging, a practice mirrored in Linux device tree security guidelines. More innovatively, their ROS 2 DDS configuration enables authentication and encryption plugins, forcing all nodes to mutual TLS—a configuration that would make any software dev agency building medical IoT devices take notes. As one cybersecurity researcher noted:

“What FRC teams are doing with DDS-Security on Jetson is essentially a dry run for securing ROS 2 in autonomous warehouse robots. If they can get mutual auth working over spotty Wi-Fi in Houston’s convention center, there’s no excuse for factories skipping it.”

The implementation mandate: here’s how you’d verify their DDS-Security setup in practice—a one-liner to check if the perception node is enforcing encryption:

ros2 topic info /ibots/vision/detections | grep -i "dds.security"

Expected output if properly configured: DDS Security: Enabled (Plugins: Builtin.Authentication, Builtin.Crypto). No output? You’re running plain DDS over UDP—inviting spoofed detection floods that could make the robot chase phantoms into the barrier. This level of observability is table stakes for any team hoping to pass a SOC 2 Type II audit on their edge deployment.

The editorial kicker: as enterprise AI teams wrestle with securing LLM APIs and vector databases, FRC teams like IBOTS are quietly solving the harder problem—securing real-time perception-to-action loops where latency and safety are non-negotiable. The next frontier isn’t bigger models; it’s formal verification of the control policy under perception uncertainty, a niche where cybersecurity auditors with controls engineering backgrounds will soon command premium rates. Watch for these kids to show up at DEF CON’s AI Village next year—not as curiosities, but as the people who actually stress-tested the stack.

*Disclaimer: The technical analyses and security protocols detailed in this article are for informational purposes only. Always consult with certified IT and cybersecurity professionals before altering enterprise networks or handling sensitive data.*