Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Unlocking Autonomous Construction Sites
Transcript
- Lucas: So you've probably seen those videos of autonomous excavators moving dirt in a perfectly choreographed loop. Looks impressive, but the real hard part isn't the machine learning — it's the latency. Luna: Wait, isn't that exactly the kind of problem cloud computing was supposed to solve? Just stream the sensor data up to a server farm and send commands back down? Lucas: That's what a lot of construction firms assumed. Then they ran the numbers. A typical excavator on a job site is generating maybe 40 gigabytes of lidar and stereo camera data per hour of operation. Sending that to the cloud — even a reasonably close region like us-west-2 — adds a 150-millisecond round trip on a good day. Luna: And that's too slow for something like collision avoidance or precise grading. Lucas: Exactly. If a dump truck is backing toward an excavator at walking speed — maybe 3 miles per hour — 150 milliseconds translates to about 8 inches of movement. That's the difference between a safe stop and a fender bender. For grading you need sub-inch accuracy. So the industry is quietly pivoting to on-site edge compute racks. Luna: I want to hear about a real deployment. Is anyone actually doing this at scale, or is it still all press releases? Lucas: The most concrete example I've seen — and I visited this site last month — is a sand and gravel quarry outside Phoenix, Arizona. It's run by a mid-sized aggregates company. They partnered with a startup called Edgeworx and a heavy-equipment OEM to create what they call a 'fully autonomous earthmoving zone' covering about 12 acres. Luna: And they have edge compute onsite? Not just a 5G connection to a regional data center? Lucas: They have a weatherized server cabinet — think a well-sealed Pelican case on steroids — bolted to a concrete pad next to the weigh station. Inside is a single NVIDIA A100 GPU, 256 gigs of RAM, and a local 5G private network base station from Ericsson. The whole thing pulls about 700 watts. They measured the round-trip inference latency at 8 milliseconds consistently. Luna: Eight milliseconds. I mean, that's basically real time for anything moving at quarry speeds. Lucas: Right. And they're coordinating four vehicles simultaneously — two autonomous excavators and two autonomous articulated dump trucks. All the sensor fusion, path planning, and collision avoidance happens on that single edge node. The cloud is only used for telemetry logging and over-the-air model updates, which happen overnight. Luna: What's the productivity story? I assume the OEM is claiming some huge efficiency gain. Lucas: They published a white paper — I've got it here — showing a 27 percent increase in tons moved per hour over the first 1,000 operating hours, compared to human-operated equipment on the same site. But the more interesting number to me is the downtime reduction. The autonomous system logged 98.3 percent operational availability. The human shift averaged 82 percent because of breaks, shift changes, and incidents. Luna: Okay, but that's a controlled quarry. What happens when you take this to a messy urban demolition site with dust, unpredictable obstacles, and pedestrians wandering by? Lucas: That is the open question. The quarry environment is almost ideal — wide open, known obstacles, no civilians. The edge hardware itself is actually pretty robust. The real challenge is the perception stack. Dust is brutal for lidar, and vibration can loosen connectors. Edgeworx told me they had to redesign their mounting brackets three times because the first two versions cracked under the constant jostling. Luna: So it's not just software — it's mechanical engineering. That's not something the typical cloud company thinks about. Lucas: Right. And that's why you're seeing partnerships between edge compute startups and traditional construction equipment manufacturers. The OEMs understand the physical constraints. The tech companies understand the AI. Neither can do it alone. Luna: I'm curious about the business model. Does the quarry own the edge hardware, or is it a service? Lucas: It's a managed service. Edgeworx provides the cabinet, the GPU, the 5G radio, and the software stack, all for a monthly fee per vehicle. The quarry pays nothing upfront. The OEM handles maintenance of the vehicles themselves. I think that's the only way this scales — construction companies don't want to manage servers. They want to move dirt. Luna: And what about the 5G piece? Are they using licensed spectrum or CBRS? Lucas: They're using CBRS — the Citizens Broadband Radio Service — in the 3.5 gigahertz band. It's unlicensed-ish, but they have a spectrum access system that coordinates with incumbents. The base station covers the entire 12-acre zone with no dead spots. Latency is under two milliseconds on the radio link itself. The bottleneck is the GPU inference, not the network. Luna: Before we go deeper, I just want to say something about this show. We deliberately don't run ads on these episodes. It's a choice we made a while ago because we wanted the conversation to be entirely driven by what's interesting, not by sponsor copy. Lucas: Yeah, it's a luxury, honestly. And it only works because some listeners decide to support that choice. So if today's tech conversation gave you something usable — a number, a case study, a way to think about edge compute — and you want to help keep it ad-free, the link is buy me a coffee dot com slash fexingo. Luna: It's a small gesture that makes a big difference for us. And it means we don't have to interrupt the flow with a spot. Lucas: Okay, back to the site. One of the less obvious benefits they're seeing is safety. The quarry had two serious incidents in the three years before automation. Since the autonomous zone went live, zero. Luna: Is that because they removed people from the danger zone, or because the edge system is better at avoiding hazards? Lucas: Both, but mostly the second. The autonomous vehicles have 360-degree lidar coverage and a 200-millisecond planning horizon. Human operators have blind spots and reaction times closer to 500 milliseconds. The edge node can also trigger a global emergency stop across all vehicles in under 10 milliseconds if any sensor detects an anomaly. That's physically impossible to do with a human in the loop. Luna: So the edge becomes a safety-critical system. That raises the bar for reliability. Lucas: Absolutely. The cabinet has redundant power supplies, redundant network interfaces, and a watchdog timer that reboots the GPU if it hangs. They also have a degraded mode where the vehicles can complete their current task and then park safely if the edge node goes dark. The cloud backup is too slow for real-time control, but it can handle graceful shutdown. Luna: What about latency to the cloud backup? Even a graceful shutdown needs some coordination. Lucas: The graceful shutdown is actually a local fallback — the vehicles have a pre-computed set of 'return to base' paths stored onboard. The edge node sends a heartbeat every 100 milliseconds. If three heartbeats are missed, each vehicle independently executes its emergency park sequence. That's all onboard. No cloud needed. Luna: That's smart. So the edge is the primary brain, but the vehicles have a simple reflex arc as a backup. Reminds me of a cockroach — you know, the central nervous system handles most behavior, but the legs have local reflexes. Lucas: Exactly. And that analogy actually maps well to the network architecture. The edge node handles global coordination — route planning, traffic management, task assignment. The individual vehicle controllers handle low-level motion, suspension, and immediate obstacle avoidance. It's a hierarchical system. Luna: So where does this go next? Is the plan to expand the autonomous zone to the entire quarry? Lucas: That's phase two, expected later this year. But the more ambitious goal is to take the same edge stack into highway construction and eventually demolition. Demolition is much harder because the environment changes every hour. The edge node would need to ingest real-time aerial drone surveys and update the digital twin continuously. Luna: That's a massive compute load. Can a single GPU handle that? Lucas: Probably not. Edgeworx is already testing a two-node cluster with a second A100 for the digital twin updates. But that doubles the power draw and the thermal management challenge. In a quarry you have open air and a generator nearby. On a demolition site you might be plugged into a temporary power pole. It's not trivial. Luna: Speaking of power, what about edge compute in remote areas where there's no grid at all? Some of the largest construction projects are in the middle of nowhere. Lucas: That's a real constraint. The Phoenix quarry has grid power, but they also have a diesel generator as backup. For truly off-grid sites, you'd need solar plus battery, and that changes the economics significantly. A 700-watt server cabinet 24/7 requires about 17 kilowatt-hours per day. That's roughly a 5-kilowatt solar array and a 20-kilowatt-hour battery bank. Doable, but adds maybe $15,000 to the upfront cost. Luna: Still small relative to the cost of the heavy equipment. One autonomous dump truck is what, half a million dollars? Lucas: About that. So the edge infrastructure is a rounding error. The real barrier is proving that the system works in unstructured environments. Once you have that proof, the capital expenditure is negligible. Luna: I want to go back to the 27 percent productivity gain. How much of that is from the autonomy itself versus just having a coordinated fleet with no breaks? Lucas: The white paper breaks it down: 12 percent from 24/7 operation — they run two 10-hour shifts with a 4-hour maintenance window — and 15 percent from optimized path planning and reduced idle time. Human operators tend to wait more, especially during shift changes and refueling. The autonomous system coordinates refueling so that only one vehicle is offline at a time. Luna: That's interesting. The edge node is essentially acting as a real-time logistics optimizer. Not just a reactive controller but a planner. Lucas: Right. And that's where the edge shines over the cloud. The optimizer runs a rolling horizon every 5 seconds, looking 30 seconds ahead. If you tried to do that over the cloud, the round-trip latency would make the plan stale by the time it arrived. The edge can iterate fast enough to keep all four vehicles moving efficiently. Luna: So the edge is less about 'moving compute closer' and more about enabling a new control loop that simply can't exist over a wide-area network. Lucas: That's exactly the argument. It's not just a performance improvement — it's a qualitative difference in what you can do. You can have a tightly coupled multi-agent system with sub-10-millisecond coordination. That's not possible with today's cloud infrastructure, no matter how good the connection is. Luna: So when do we see this on a real highway construction project? Any timelines? Lucas: I've heard off the record that a major US highway contractor is planning a pilot for late 2026 or early 2027. They want to automate the paving train — the convoy of dump trucks, pavers, and rollers. The edge node would coordinate the entire train's speed and spacing to produce a more uniform road surface. That's a huge quality improvement if it works. Luna: I can see the logic. A smoother road lasts longer, which saves taxpayer money. But the safety case for a public road with live traffic is much tougher than a quarry. Lucas: Absolutely. The quarry is in many ways the perfect sandbox — literally. The next step will be proving the system can handle mixed traffic, pedestrians, and unpredictable obstacles. That's why I think the first highway deployments will be at night or in closed lanes. You need to de-risk before you mix autonomous and human drivers. Luna: It feels like we're at a tipping point. The hardware is ready, the software is getting there, and the business model is proven in a low-risk environment. The question is whether the industry has the appetite to push into more complex sites. Lucas: And that's where the edge compute providers have a role to play. They need to make the system as easy to deploy as a portable generator. Plug it in, turn it on, and it just works. If they can do that, I think we'll see autonomous construction spread faster than most people expect. Luna: And if they can't, it'll remain a fascinating pilot that never leaves the quarry. Lucas: Right. The next 18 months will tell us which way it goes. But the early data is promising — 27 percent productivity, zero safety incidents, 98 percent uptime. That's hard to ignore. Luna: Especially in an industry that's famously slow to adopt new technology.