Benchmark Drops With Chatgpt And Camera

Yes - what you’re seeing is very typical when camera streaming + frequent EZ-B reads + cloud TTS/ChatGPT are all competing on the same ARC project path.

Why it happens

A few things are stacking up:

• Benchmark / Read ADC is a frequent blocking read from the EZ-B
• Camera streaming adds continuous bandwidth and processing overhead
• ChatGPT + TTS adds network calls and audio playback work
• On Wi-Fi, all of those share the same radio/network path and can create latency spikes
• When latency spikes, Movement Panel feels laggy and audio stutters, which matches your symptoms

Even if CPU is low, the bottleneck is usually:

• EZ-B communication latency
• Wi-Fi contention
• blocking reads/timeouts
• camera/audio/network activity interfering with one another

What I’d recommend

1) Keep the camera resolution as low as practical

You already found this helps.

For conversational driving, I’d try:

• 160x120
• 5 fps
• lowest acceptable compression
• only enable tracking if you actually need it

ARC’s camera processing gets expensive fast. The Camera Device skill also notes that higher resolutions can heavily impact performance.

2) Don’t use Benchmark as a "live" performance metric while everything is active

Benchmark is useful for testing, but it itself contributes to traffic.
If you want to measure real performance under load, use it briefly before/after, not continuously.

3) Prefer a topology where the EZ-B has the cleanest Wi-Fi path possible

For your use case, the most stable setups are usually:

• Laptop on Ethernet
• EZ-B on a dedicated 2.4 GHz Wi-Fi network
• No extra Wi-Fi client chatter on the robot network
• If internet is needed, let the laptop do the internet side, not the EZ-B

Your idea of Ethernet on the laptop + separate 2.4 GHz USB Wi-Fi for the EZ-B is generally the right direction.

If AP mode is stable but internet access is the issue, a common approach is:

• Use laptop Ethernet for internet
• Use Wi-Fi adapter #1 for your home router
• Use Wi-Fi adapter #2 or routing/ICS for the EZ-B link
• Keep the EZ-B network isolated as much as possible from other traffic

4) Reduce how often scripts poll the EZ-B

If any scripts are reading ADC, ping sensors, or other ports too frequently, they can flood the communication channel.

The ARC performance guide calls out:

• tight loops
• flooded EZ-B communication
• multiple skills querying the same port
• slow sensor timeouts

If you have any custom scripts, make sure they include sleep() and aren’t hammering the same ports repeatedly.

Example:

while (true) {
  var adcVal = ADC.get(ADC0);
  print(adcVal);
  sleep(100);
}

5) Minimize overlapping reads

Make sure you don’t have:

• a Benchmark skill
• a Read ADC skill
• a script polling ADC0
• another skill using the same sensor/port

all at once unless you truly need them. Multiple consumers of the same port can absolutely lower throughput.

6) Consider splitting responsibilities

For a project like this, it can help to separate tasks:

• EZ-B #1: movement + immediate control
• EZ-B #2: camera or sensors, if applicable
• Or keep camera on the laptop side and minimize robot-side polling

The goal is to avoid the robot’s primary control link doing everything at once.

About ChatGPT specifically

The ChatGPT skill itself usually isn’t the main problem - but the timing of requests, TTS playback, and any logic that waits on responses can increase lag when combined with camera activity.

If your ChatGPT replies trigger audio, that can further add perceived delay when the system is already under communication load.

Practical target

If your goal is above ~200 reads/sec while camera + ChatGPT are active, I’d aim for:

• camera at 160x120
• low fps
• no extra polling scripts
• no redundant skills reading the same ports
• laptop on Ethernet
• EZ-B on a dedicated Wi-Fi connection
• avoid AP mode if it breaks your internet workflow, unless you can bridge/ICS cleanly

Quick diagnostic question

Can you tell me:

1. Are you using any scripts that poll ADC0 or movement continuously?
2. Is the Camera skill just streaming, or is it also doing tracking?
3. Does ChatGPT speak through PC TTS or the EZ-B audio?
4. Roughly how many ARC skills are active when the slowdown happens?

If you want, I can help you design the best ARC topology for conversational driving so ChatGPT, camera, and movement all coexist more smoothly.