Battle Motors Cuts Truck Fires and Contamination Fines With Edge AI

Presentation by Paul Marsolan, Battle Motors — Ai4

Ai4

Battle Motors uses native edge computing on heavy refuse trucks to prevent catastrophic hopper fires, audit waste contamination in real time, and eliminate cellular dependency.

AI-assisted summary of Presentation by Paul Marsolan, Battle Motors — original source. Check the original for full context.

Source published: 2026-08-23

Detailed notes

## Background This presentation was delivered by Paul Marsolan, Chief Software Officer at Battle Motors, and published on YouTube by the Ai4 channel on August 23, 2026. According to the page description, the talk provides an insider look at Fortris™, Battle Motors' edge vision platform designed for 80,000-pound refuse trucks operating in harsh commercial environments.

## TL;DR Paul Marsolan explains how Battle Motors

## Key takeaways - [00:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=4s) Integrating edge AI and modern software into vocational refuse trucks optimizes route efficiency, prevents contamination, and mitigates catastrophic vehicle fires. - [01:45](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=105s) Battle Motors replaces fragmented third-party aftermarket hardware with a unified, factory-integrated edge compute platform delivering sub-200 ms latency. - [04:14](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=254s) All critical detection models run completely offline on edge hardware to ensure uninterrupted safety monitoring in cellular dead zones. - [05:58](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=358s) The Fortris Hub combines automotive-grade dual digital displays with mTLS-encrypted cloud telemetry for hardware-verified security. - [07:42](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=462s) The system runs parallel Nvidia-powered vision models filtered by mechanical action triggers to optimize compute efficiency during active collection. - [09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) Early fire detection uses multi-frame verification and synthetic training data to prevent total loss of $300k+ assets, supported by driver-in-the-loop feedback. - [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) Multi-modal sensing automates proof-of-service verification, identifies bin overfill for variable billing, and logs municipal road hazards like potholes. - [17:11](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1031s) An automated cloud AI pipeline using VLMs and LLMs auto-annotates edge-captured images to continuously update models via OTA deployment. - [17:49](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1069s) The Fortris platform drastically simplifies cab ergonomics by replacing 3–4 cluttered aftermarket screens with two integrated OEM displays. - [18:49](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1129s) Fire detection models intentionally lean toward false positives because missing an active vehicle fire carries catastrophic risk. - [20:52](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1252s) Privacy protocols automatically blur faces and license plates during ingestion and enforce a 14-day deletion rule for raw image data. - [22:10](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1330s) Local on-chip execution was chosen over cloud inference to eliminate cloud operational costs and prevent critical alert delays in poor LTE coverage areas. - [23:19](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1399s) Battle Motors monetizes the Fortris platform via upfront hardware purchases, software subscriptions, or bundled Hardware-as-a-Service (HaaS) options.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=4s) Rethinking the Garbage Truck as an AI Platform

Every day, people generate waste and watch garbage trucks stop by their homes to empty bins, giving little thought to where the trash goes once it disappears. While this invisible process keeps society running smoothly, there is a massive opportunity to leverage artificial intelligence on these commercial fleets. Integrating AI into refuse vehicles can radically optimize route efficiency and improve safety for haulers, equipment, and surrounding pedestrians.

Battle Motors is the second largest vocational truck OEM in the United States, producing custom commercial vehicles out of a 325,000-square-foot facility in New Philadelphia, Ohio. As Chief Software Officer, Paul Marsolan focuses on solving complex operational challenges using modern software architecture. Key problems include stopping recycling stream contamination across refuse, recycling, and compost routes, as well as mitigating catastrophic vehicle fires. Historically, a driver only learned their truck was ablaze when another motorist flagged them down at a red light—often resulting in the total destruction of a multi-hundred-thousand-dollar asset. By deploying edge AI vision systems natively into the vehicle, Battle Motors can detect emerging hazards instantly, protect pedestrians, and manage bin contamination at the point of collection.

- **Manufacturer Scale:** Battle Motors operates a 325,000 sq ft manufacturing facility in New Philadelphia, Ohio. - **Industry Standing:** Second largest vocational truck OEM in the United States. - **Vehicle Weight:** Heavy vocational trucks operate at gross vehicle weights up to 80,000 lbs.

---

[01:45](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=105s) Consolidating Legacy Aftermarket Hardware into Native Edge Architecture

The commercial refuse industry traditionally relies on a highly fragmented supply chain. Battle Motors builds and ships a cab-and-chassis to third-party body builders, who mount the rear packer or hopper assembly. Finally, local end-customers (such as the City of Los Angeles) install aftermarket vision systems, route-management tablets, and driver monitors. This piecemeal approach clutter the cab with disparate screens and complex wiring, creating driver distraction.

Battle Motors replaced this disconnected setup with a unified factory-integrated hardware and software platform. Drawing inspiration from consumer electric vehicles, the platform centralizes camera feeds onto an integrated dashboard and processes vision data locally via an onboard GPU. This local edge compute architecture delivers sub-200 millisecond response times—critical for real-time driver alerts and automated intervention when pedestrians or hazardous items are detected.

``` +-----------------------------------------------------------------+ | Battle Motors Cab & Chassis | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | 8x RGB Edge Vision Camera Array | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Fortris™ Centralized Platform & Onboard Nvidia GPU | | - Local Model Execution (<200 ms latency) | | - Real-Time ADAS & Safety Alerts | | - 360-Degree Surround-View Display | +-----------------------------------------------------------------+ | (Cellular Link) v +-----------------------------------------------------------------+ | Fortris™ Cloud Command Server | | - Fleet Analytics & Over-the-Air Model Updates | | - Driver Coaching & Continuous Model Training | +-----------------------------------------------------------------+ ```

- **Camera Array:** Eight vision cameras wrap around the vehicle to eliminate blind spots. - **Latency Requirement:** Real-time inferencing occurs locally in under 200 ms to enable rapid automated responses. - **Standardized Experience:** Brings modern consumer EV conveniences—such as high-definition 360-degree reverse views—to heavy 80,000-pound commercial trucks.

---

[04:14](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=254s) Edge Compute Independence and Contamination Fines

Garbage trucks operate across diverse municipal environments, frequently encountering cellular dead zones in rural pockets, urban canyons, or dense alleyways. To ensure reliable operations without cloud dependency, all critical detection models run fully on edge hardware inside the truck cab.

Local edge processing enables instant recognition of dangerous items before they enter the compaction sequence. When an automated side-loader dumps a bin, the vision system scans for volatile hazards like propane tanks, lithium-ion batteries, and laptop electronics, warning the driver not to engage the packer blade.

Additionally, the platform addresses regulatory compliance. In states like California, haulers face steep financial penalties if municipal waste streams exceed maximum contamination thresholds (e.g., rigid plastics or trash bags mixed into organic compost or clean paper recycling). While low levels of contamination (around 20% to 30%) are often tolerated by processing facilities, exceeding these limits triggers escalating fines. Haulers are increasingly using vision logs to track non-compliant disposal back to specific residential accounts, issuing warnings or passing fines directly to offending customers.

- **Compaction Hazards:** Detecting propane tanks, car batteries, laptops, and cellular devices before compaction prevents explosions and chemical fires. - **Regulatory Baselines:** Contamination allowances typical hover between 20% and 30%; exceeding these limits results in direct municipal fines on the hauler. - **Offline Operations:** Local model inferencing guarantees full detection capabilities regardless of active LTE or cellular connectivity.

---

[05:58](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=358s) Automotive-Grade Hardware and Security Integration: The Fortris™ Hub

Commercial vocational cabs require industrial-grade electronics capable of enduring severe vibration, heat, and dust. Battle Motors designed its proprietary automotive-grade edge compute hardware, known as **Fortris Hub**, from the ground up to serve as the core vehicle intelligence center.

``` +-----------------------------------------------------------------+ | DRIVER INTERFACE | | +---------------------------+ +---------------------------+ | | | 10.2" Instrument Cluster | | 14.6" Center Console Display| | | | (Primary Gauges/Alerts) | | (360 Cam, Maps, Telematics)| | | +---------------------------+ +---------------------------+ | +-----------------------------------------------------------------+ | (mTLS Encrypted Channel) v +-----------------------------------------------------------------+ | Fortris™ Cloud Command Server | | - Real-Time Fleet Tracking & Route Progress Monitoring | | - Address-Level Pickup Verification & Contamination Logs | +-----------------------------------------------------------------+ ```

The cabin architecture features dual digital displays: a 10.2-inch digital instrument cluster directly in front of the driver and a 14.6-inch central touch screen, delivering a user experience reminiscent of a Tesla Model X. Telemetry, video clips, and event logs stream back to the central Fortris Command Server via mutual TLS (mTLS) encryption, ensuring hardware-verified security that surpasses typical commercial banking applications. Fleet operators monitor routes in real time, validating address-level service events—such as confirming why a driver bypassed a bin or identifying specific hazard items disposed of at a given location.

- **Display Configuration:** Features a 10.2-inch digital driver instrument cluster alongside a 14.6-inch central touch display. - **Security Standard:** Employs mutual TLS (mTLS) authentication for secure, encrypted communications between vehicle hardware and cloud endpoints. - **Analog Replacement:** Replaces traditional analog switches and standalone third-party monitors with a single, factory-installed digital environment.

---

[07:42](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=462s) Parallel Model Architecture and Operational Mission Filtering

To handle complex real-time video streams from eight cameras, Fortris executes multiple specialized vision models simultaneously on an onboard Nvidia chip:

1. **Generalist Model:** Scans the vehicle's surrounding environment for general dynamic objects, including pedestrians, cyclists, dogs, and cats. This model addresses severe safety risks, such as individuals sleeping inside commercial dumpsters during cold weather in cities like New York, preventing tragic incidents when bins are dumped into the hopper. 2. **Specialized Refuse Model:** Focuses on the hopper interior and collection zones to classify specific waste materials, distinguishing between glass, paper, plastic, and high-risk hazards like car batteries or propane tanks. 3. **Action-Detection Model:** Monitors mechanical components, such as the lift arm, to infer the truck's precise operational state.

Rather than running heavy classification continuously, system logic uses the action-detection model as a trigger filter. The vision system records a continuous background DVR for general collision liability, but localized detection models activate full classification only when the lift arm engages. Filtering events by mechanical action dramatically reduces unnecessary compute overhead and keeps model inferencing focused on the active loading sequence.

- **Compute Hardware:** Powered by high-performance onboard Nvidia embedded GPUs (such as the Jetson Orin series). - **Transient Risk Factors:** Specialized object recognition flags human presence in commercial dumpsters prior to mechanical compaction. - **Selective Data Processing:** Focuses intense model inferencing precisely on operational cycles rather than ambient driving conditions.

---

[09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) Mitigating Hopper Fires via Synthetic Data and Driver Feedback

Hopper fires represent one of the most costly risks in refuse collection, typically progressing through three distinct severity levels:

``` +-----------------------------------------------------------------+ | FIRE ESCALATION TIER LIST | | | | [Level 1] Early Incipient Fire / Smoke | | -> Action: In-cab alert, pull over, follow checklist | | | | [Level 2] Active Load Combustion | | -> Action: Eject burning load in safe open area | | | | [Level 3] Total Vehicle Loss | | -> Result: Complete destruction of $300k+ asset | +-----------------------------------------------------------------+ ```

To catch fires at Level 1, Battle Motors trained specialized vision models using synthetic imagery—overlaying photorealistic smoke and fire elements onto real hopper base photos to teach the system to recognize incipient ignition.

When the model detects smoke or fire, it triggers an instant audio-visual warning on the main display, providing the operator with fleet-approved emergency protocols. Because environmental conditions like dust from construction debris or brightly colored orange bags can occasionally mimic fire or smoke, the system requires multi-frame verification before issuing an alert. Operators can confirm or dismiss alerts directly on the touchscreen. This driver feedback loops straight back into the cloud dataset, continually refining model precision and eliminating false positives over time.

- **Alert Escalation:** Gives drivers immediate step-by-step operational checklists to safely manage incipient fires. - **False Positive Filtering:** Analyzes sustained frame sequences to distinguish airborne dust and brightly colored debris from actual smoke or flames. - **Active Learning Loop:** In-cab driver confirmations ("Fire" or "False Alarm") automatically label edge cases for retrain cycles.

---

[13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) Customer Service Verification, Overfill Detection, and Smart City Telematics

Disputes between residential customers and municipal haulers frequently arise when homeowners claim their trash was missed, while drivers maintain that no bin was placed at the curb. The Fortris platform solves this by pairing GPS routing data with visual evidence. When a truck passes a house without servicing a bin, the system flags the route coordinate and archives a snapshot showing the empty curb. Haulers can automatically email this video proof to residents, eliminating contentious service calls.

``` +-----------------------------------------------------------------+ | MUNICIPAL DATA FLYWHEEL | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Daily Refuse Routes (Every Street / Every Week) | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Multi-Modal Edge Sensing | | - Proof of Service Verification | | - Residential/Commercial Bin Overfill Detection | | - Automated Pothole & Road Hazard Identification | | - Damaged Traffic Sign & Signal Outage Detection | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Municipal API Integration & Public Works | +-----------------------------------------------------------------+ ```

The system also targets commercial and residential bin overfill. Haulers lose revenue when servicing overflowing dumpsters, and cities are increasingly auditing residential bins for lids pushed open by excess waste. Battle Motors is actively training models on consumer-grade trash cans to automatically detect overfilled bins and append surcharge line-items to customer accounts.

Beyond waste management, refuse trucks act as continuous mobile sensor platforms for smart city administration. Because garbage trucks traverse almost every municipal street on a predictable weekly schedule, their outward-facing cameras can monitor public infrastructure. Battle Motors works with municipal partners to train edge models that automatically log road hazards—identifying potholes, non-functioning traffic signals, and missing stop signs—and route maintenance tickets directly to city public works departments.

- **Proof of Service:** Automatically captures geo-tagged image evidence when a bin is missing, eliminating "missed pickup" disputes. - **Overfill Auditing:** Scans residential and commercial bins for illegal overfilling to automate variable-rate billing. - **Infrastructure Auditing:** Transforms municipal refuse fleets into automated road-inspection networks for public works management.

---

[17:11](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1031s) Automated Training Pipelines and Multimodal Dataset Curation

Building reliable edge models for harsh environments requires massive amounts of high-quality training data. Raw image feeds captured by the fleet stream into an automated cloud AI pipeline for processing and curation.

``` +-----------------------------------------------------------------+ | Raw Vehicle Images & Telemetry | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Automated Cloud Pipeline | | - Multimodal LLM/VLM Auto-Bounding & Bounding Box Generation | | - Automated Labeling & Synthetic Data Integration | | - Human-in-the-Loop Edge-Case Curation | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Retrained Model Deployment (OTA) | +-----------------------------------------------------------------+ ```

To eliminate manual annotation bottlenecks, Battle Motors leverages vision-language models (VLMs) and large language models (LLMs) to automatically detect objects, draw bounding boxes, and label incoming video frames. While software leadership still handles targeted manual curation for complex edge cases, automated labeling performs the vast majority of dataset annotation. This automated pipeline rapidly turns daily operational logs into refreshed, highly accurate model weights, which are deployed back to the fleet via over-the-air (OTA) updates.

- **Labeling Automation:** Uses multimodal AI models to automatically bound and annotate raw frame uploads, replacing manual third-party labeling services. - **Continuous Deployment:** Edge dataset additions directly feed retrained vision models back to fleet vehicles via secure cloud OTA channels.

Here is a detailed, faithful walkthrough of the second part of the presentation:

[17:43](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1063s) Data Pipeline, ADAS Integration, and Cab Simplification

When training models on new edge data, incoming images are systematically compared against the existing on-truck model to evaluate performance improvements using training and validation tests. Because a typical refuse truck completes roughly 1,000 stops per day—visiting about 1,000 individual houses—the system processes and evaluates thousands of images per truck daily.

Beyond standard vision monitoring, Battle Motors integrates safety-critical Advanced Driver Assistance Systems (ADAS) features into the vehicle platform. Bringing safety capabilities like automatic braking into these heavy-duty trucks offers significant benefits to haulers and municipalities. Combining ADAS with driver coaching, digital video recording (DVR) capabilities for accident liability protection, and edge vision creates a comprehensive safety and operational ecosystem.

``` [ 8x RGB Cameras ] ---> [ Edge AI (Jetson Orin) ] | +-----------------------------------+-----------------------------------+ | | | v v v [ Real-Time Safety ] [ Integrated ADAS ] [ Data Pipeline ] (Fire Detection) (Automatic Braking) (1k stops/day data) | | +-----------------+-----------------+ | v [ Fortris™ Dual Screens ] (Replaces 4+ cluttered) ```

As a primary Original Equipment Manufacturer (OEM), Battle Motors enjoys a unique advantage in cab design. A traditional refuse truck interior is cluttered with a chaotic array of three to four separate aftermarket monitors—one tablet for route navigation, a dedicated screen displaying camera feeds, and additional monitors for pickup task lists. In contrast, the factory-installed Fortris™ Hub reduces this clutter down to just two streamlined screens. These dual displays consolidate all critical operational context directly in front of the driver straight off the factory line, replacing piecemeal aftermarket add-ons.

- Typical refuse trucks visit approximately 1,000 homes or stops per shift, yielding thousands of training images per truck every day. - The Fortris™ Hub replaces 3–4 separate cab screens with 2 integrated OEM displays.

[17:49](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1069s) Q&A: False Positives in Fire Detection

During the audience Q&A session, a member of the audience asks about the acceptable false positive rate given the platform's bias toward false alarms. Paul Marsolan explains that for high-stakes hazards like hopper fires, the detection model intentionally leans toward triggering false positives. A fire in an 80,000-pound refuse truck is a catastrophic event, so it is far better to alert the operator—allowing them to visually verify the situation—than to risk a missed detection.

While truck fires are infrequent enough that precise percentage figures are not highlighted, they occur often enough to remain a top concern for fleet owners and municipalities. The system prioritizes flagging potential fire indicators early while providing operators with sufficient context on-screen to quickly verify the alert.

[20:52](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1252s) Q&A: Camera Placement and Fleet Privacy Protocols

Addressing a question regarding visual hardware placement, Marsolan notes that camera mounting depends on the specific vehicle chassis configuration and operational objective. For example, for customer pickup verification, cameras are typically installed on the radiator area behind the cab on Battle Motors chassis, or directly on the body for other vehicle types. Blind-spot and side-view cameras are mounted on the sides of the cab rather than on the side mirrors.

In response to concerns about capturing imagery on private property, Marsolan clarifies that privacy management is built directly into the data pipeline. Mirroring consumer mapping platforms like Google Street Maps, the pipeline automatically blurs faces and license plates as imagery is ingested. To further safeguard data integrity and privacy, raw images are automatically deleted after approximately 14 days.

- Privacy measures automatically blur faces and license plates in recorded video. - All raw edge image data is routinely discarded/deleted after a 14-day retention period.

[22:10](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1330s) Q&A: On-Chip Edge Computing vs. Cloud Inference

When asked why Battle Motors chose local on-chip execution rather than cloud-based LLM inference, Marsolan shares that the company originally explored cloud inference. However, two major factors rendered a pure cloud architecture impractical: cloud processing costs and cellular connectivity limits. Large language models and heavy cloud processing introduce substantial recurring expenses without offering proportional real-time value on the vehicle.

More critically, many municipal and private fleets operate in rural areas lacking reliable LTE or T-Mobile cell service. Delayed alerts are non-viable when immediate safety is at stake: "I can't be like, 'Hey, there's a fire in your truck, but only when you get back to T-Mobile cell phone service.' That doesn't really work out very well." While non-real-time workloads could theoretically leverage cloud models, localized GPU processing keeps immediate safety alerts instant and keeps operational costs low. On-chip processing also enables proactive customer outreach—such as notifying residents about overfilled cans or missed pickups before citizens call in to complain.

[23:19](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1399s) Q&A: Customer Behavior, Policy Approval, and Commercial Pricing

Responding to a query on whether automated driver/fleet feedback has changed customer waste habits, Marsolan notes that no statistically significant change in citizen behavior has been observed yet. This is primarily because participating cities have not fully rolled out public education programs. Many municipal leaders and city councils are still working through privacy policies and administrative approvals before formally deploying automated feedback loops to residents.

Regarding commercial pricing and deployment models, Battle Motors offers flexible subscription structures: - **Hardware Acquisition:** Customers can purchase hardware pre-installed directly from the factory on new Battle Trucks or retrofit existing fleets via aftermarket kits. - **Software & Services:** Platform features are delivered through a cloud/edge software subscription. - **Hardware-as-a-Service (HaaS):** Equipment costs can either be paid upfront or bundled alongside software into a single monthly SaaS/HaaS recurring fee per truck.

## Notable quotes > [22:10](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1330s) I can't be like, 'Hey, there's a fire in your truck, but only when you get back to T-Mobile cell phone service.' That doesn't really work out very well. — Paul Marsolan

## Mentioned Paul Marsolan, Battle Motors, Fortris, Fortris Hub, Nvidia, Jetson Orin Nano, City of Los Angeles, New York, Tesla Model X, Google Street Maps, LTE, T-Mobile, Ai4

## Video description Paul Marsolan - Chief Software Officer, Battle Motors

What does it take to ship production AI on an 80,000-pound refuse truck — a vehicle that fights vibration, heat, dust, and intermittent LTE on every shift, and where a single missed detection can mean a hopper fire from a discarded lithium battery?

This talk goes inside Fortris™, Battle Motors' edge vision platform: eight RGB cameras, three vision models running in parallel on a Jetson Orin Nano, sub-200 ms decision latency, and a bidirectional cloud plane that turns every route into a labeled dataset. We'll walk through the real architecture — specialist + generalist + action-detection models, OTA model deployment with per-truck rollback, ADAS sensor fusion — and the production lessons behind it. Deploying a model to 1,000 trucks is not deploying to 1,000 web servers; reversibility is a product feature, not a nice-to-have.

You'll leave with a concrete picture of what edge AI looks like when it has to work in one of the harshest commercial environments in America — and why the cab of an 80,000-pound truck might be the next great human-AI interface.

[Source](https://www.youtube.com/watch?v=gdUAQLIMqxM)

中文简报

Battle Motors 借助边缘 AI 减少卡车火灾与污染罚款 Battle Motors 在重型垃圾车上采用原生边缘计算，以预防灾难性的料斗火灾、实时审核废弃物污染，并消除对蜂窝网络的依赖。 Paul Marsolan 介绍了 Battle Motors 如何通过在专业垃圾车中集成边缘 AI 与现代软件平台，实现低延迟离线运行视觉模型，从而在无需蜂窝网络连接的情况下，有效预防灾难性车辆火灾并实时监测废弃物污染。 Battle Motors 推出专为商用垃圾车打造的 Fortris 原厂集成边缘计算平台，取代传统碎片化后装设备，通过车载 Nvidia GPU 实现低于 200 毫秒的本地推理延迟。 (1:45) 为解决偏远或密集作业区域的网络盲区问题，关键检测模型完全在边缘离线运行，可在压实前识别锂电池与丙烷罐等危险品，防止火灾并降低违规混杂罚款。 (4:14) 系统在边缘端并行执行通用环境模型、垃圾专用分类模型及机械动作模型，通过机械臂提升动作精准触发深度推理，显著降低持续运算负载。 (7:42) 平台结合合成训练数据与多帧验证实现早期料斗火灾预警，配合驾驶舱人机协作确认机制，有效避免价值 30 万美元以上重型资产的彻底损毁。 (9:55) 环卫车辆借助多模态感知化身移动基础设施巡检平台，不仅能自动生成服务凭证与垃圾溢出账单，还能向市政部门上报路面坑洞和受损路标。 (13:04)

Reading summary

Paul Marsolan, Chief Software Officer at Battle Motors, argues that replacing fragmented aftermarket hardware with factory-integrated edge AI on refuse trucks enables sub-200 millisecond hazard intervention, curbs costly waste contamination fines, and converts collection fleets into smart city sensor platforms. Integrated Edge Architecture and Offline Vehicle Compute (1:45) Battle Motors, the second largest vocational truck OEM in the United States, designed its proprietary Fortris Hub to replace the fragmented aftermarket hardware typical of municipal refuse fleets. Instead of burdening drivers with three to four independent monitors, navigation tablets, and add-on camera displays, the OEM installs a consolidated dual-screen digital environment—a 10.2-inch instrument cluster and a 14.6-inch center console display—secured by mutual TLS (mTLS) encryption. To support real-time driver alerts and integrated Advanced Driver Assistance Systems (ADAS) like automatic braking on trucks weighing up to 80,000 pounds, vision models run directly on an onboard Nvidia GPU rather than relying on the cloud. Marsolan emphasized that relying on cloud inference for vehicle safety is impractical due to recurring compute costs and frequent cellular dead zones in rural routes and dense urban corridors. By executing models locally at the vehicle edge, the platform guarantees sub-200 millisecond response times, ensuring safety-critical alerts function without an active cellular connection. Preventing Compaction Hazards and Costly Hopper Fires (4:14) Commercial refuse vehicles face severe financial and safety risks from volatile items entering the compaction cycle, including propane tanks, car batteries, and consumer electronics. The Fortris platform scans waste during bin dumping, alerting drivers to halt the packer blade before chemical fires or explosions ignite. The system also tracks regulatory compliance; in jurisdictions like California, haulers incur steep municipal penalties when contamination in recycling or compost streams exceeds standard 20% to 30% allowances. Vision logs allow haulers to trace non-compliant items back to specific residential addresses to issue warnings or pass through municipal fines. Hopper fires can rapidly escalate to complete vehicle destruction, destroying assets valued at over $300,000. Battle Motors trained edge vision models on synthetic smoke and fire imagery blended with hopper base photos to catch incipient ignition at Level 1. Because the cost of a missed detection is catastrophic, Marsolan explained that the system intentionally tolerates false alarms, using multi-frame verification to filter airborne dust and relying on driver touchscreen confirmations to continually refine the active learning loop. Event-Driven Vision Processing and Automated Data Pipelines (7:42) To process feeds from an eight-camera array without exhausting edge compute capacity, the Fortris system executes specialized models in parallel using operational event triggers. A generalist model monitors the truck's surroundings for pedestrians, cyclists, and vulnerable individuals sheltering inside commercial dumpsters. However, intensive waste classification inside the hopper activates only when an action-detection model senses the mechanical lift arm engaging. This mechanical filtering limits compute overhead to active pickup events while preserving continuous DVR footage for liability reviews. Because a single refuse truck visits roughly 1,000 homes per shift, the fleet captures thousands of daily images. To bypass manual data-labeling bottlenecks, Battle Motors feeds edge uploads through an automated cloud pipeline utilizing vision-language models (VLMs) and large language models (LLMs) to generate bounding boxes and annotations. Refreshed model weights are then redeployed back to the fleet via over-the-air (OTA) updates. Service Verification, Infrastructure Telematics, and Privacy Safeguards (13:04) The platform pairs GPS routing with camera snapshots to eliminate customer disputes over missed pickups. When a driver passes an address without servicing a bin, the system flags the coordinates and archives visual proof of the empty curb to verify why service was bypassed. Battle Motors is also training models to identify overfilled residential and commercial bins, enabling automated customer surcharges. Because collection trucks navigate almost every municipal street on a weekly cadence, outward-facing cameras double as automated public works inspection tools, logging potholes, missing traffic signs, and defective signals. To address privacy concerns associated with mobile recording, Marsolan stated that the platform's ingestion pipeline automatically blurs faces and license plates, mirroring consumer mapping standards. In addition, raw edge video is automatically purged after a 14-day retention cycle. Deployments are structured flexibly, offering pre-installed factory options, aftermarket retrofit packages, and unified Hardware-as-a-Service subscriptions. While the technical platform automates service verification and hazard logging, Marsolan observed that citizen disposal habits have not yet shown statistically significant improvement, as many municipalities are still working through administrative approvals and privacy policies required to launch resident-facing education and automated enforcement programs.

精炼阅读摘要

Battle Motors首席软件官Paul Marsolan提出，通过在重型环卫卡车上原生集成边缘计算硬件与多模态AI视觉系统，不仅能化解传统后装设备割裂造成的驾驶安全隐患，还能在无网络依赖的前提下实现起火预警、废弃物合规监管与智慧城市基础设施巡检。车载端侧原生架构与座舱整合 (1:45) 传统商用环卫行业供应链高度分散，底盘制造商完成车架后交由车体改装厂加装后斗，最终由市政客户在驾驶室内拼凑各类第三方平板电脑、倒车监视器与线路，导致座舱混乱且容易分散驾驶员注意力。作为美国第二大专业卡车原始设备制造商（OEM），Battle Motors在其位于俄亥俄州纽费拉德尔菲亚占地325,000平方英尺的工厂中，针对总重达80,000磅的重型卡车推出了名为Fortris Hub的原生软硬件一体化平台。该平台效仿乘用电动车的整合设计，以10.2英寸全液晶仪表盘与14.6英寸中控触控屏取代了原本繁杂的3至4块外挂屏幕。针对车辆常在偏远区域、高楼峡谷或狭窄小巷面临蜂窝信号中断的问题，系统完全依赖车规级板载Nvidia GPU（如Jetson Orin系列）进行本地边缘推理，确保响应延迟低于200毫秒。Marsolan强调，对于高危作业场景，系统不可能等待恢复网络连接后再发送警报；本地执行不仅消除了持续调用云端大语言模型的昂贵带宽与算力成本，还通过mTLS加密通道保障了车云数据传输安全。并行视觉模型与动作事件触发机制 (7:42) 为了在消除盲区的同时高效处理来自全车8个RGB摄像头的高清视频流，Fortris平台在板载芯片上并行运行三个专精化的计算机视觉模型：一是通用环境模型，负责持续检测卡车周围的行人、骑行者与宠物，并特别针对纽约等严寒城市中流浪者在商用垃圾箱内取暖的致命风险进行识别，防止倾倒时发生意外；二是垃圾专用模型，专注于车斗内部物料识别，可区分玻璃、纸张、塑料以及易燃易爆危险物；三是机械动作检测模型，专门监控装载臂的工作状态。由于高频持续运行深度学习分类模型会消耗巨大的车载计算资源，系统采用了以机械动作为事件触发的过滤逻辑。车辆平时仅保持背景行车记录（DVR）以供事故责任追溯，只有当动作模型检测到机械臂抓取并倾倒垃圾箱的瞬间，系统才会全面激活车斗内的高负载分类推理。这种事件驱动设计极大地削减了冗余算力开支，使核心模型高度聚焦于作业周期。料斗火灾防御与合成数据闭环 (9:55) 车斗起火是环卫运营中最昂贵的风险之一，其发展通常分为三个阶段：初期冒烟起火（一级）、整车垃圾剧烈燃烧（二级，需将着火垃圾紧急倾卸至安全开阔地）以及最终整车烧毁（三级，导致单台价值超过30万美元的卡车彻底灭失）。为在初期阶段捕捉火情，Battle Motors利用合成数据技术，将逼真的烟雾与火焰效果合成到真实车斗背景中，训练出了高度敏感的火灾识别模型。系统采用多帧连续验证机制，以滤除扬尘和明亮橙色垃圾袋引发的视觉干扰。 Marsolan指出，面对80,000磅重型卡车起火的灾难性后果，火灾检测模型在设计上倾向于容忍较高的误报率，宁可触发假警报由驾驶员确认，也绝不能漏报。当系统在中控屏上弹出视听警报与标准应急处置清单时，驾驶员可直接在屏幕上确认或忽略警报。这一司机交互标签会实时回传至云端，结合多模态视觉大模型（VLM）与LLM的自动化边界标注管道，针对边缘案例持续微调模型权重，并通过OTA空中升级重新推送到车队，形成主动学习闭环。污染追溯、市政巡检与数据隐私治理 (13:04) 在加利福尼亚等州，如果混入有机堆肥或纸质回收物中的塑料或垃圾超过20%至30%的法定容许阈值，清运服务商将面临高额市政罚款。垃圾专用模型可在垃圾倒入瞬间检测出丙烷罐、汽车蓄电池、笔记本电脑等危险违禁品，并在压实刀片启动前警示驾驶员，同时记录违规投放证据，协助清运商将罚款追溯并转嫁至具体违规住户。此外，结合GPS路径与视觉凭证，系统能自动抓拍未放置垃圾桶的空置路沿，消除居民投诉漏收的争议，并自动识别超载未合盖的垃圾桶以执行额外计费。得益于环卫卡车每周全覆盖行驶每一条市政街道的特性，Battle Motors正与市政工程部门合作，将卡车转变为移动传感平台，利用朝外的摄像头自动识别道路坑洼、损坏交通标志与失效信号灯，自动生成工单直连市政公用事业系统。在隐私与商业模式方面，系统在数据入库时会自动模糊人脸与车牌，原始图像存储约14天后自动销毁；客户既可选择出厂预装或后装套件，也可采用软硬件打包的月度HaaS订阅模式。 Marsolan坦言，尽管车载边缘视觉技术已能够实现住户违规投放与超载的精准证据留存，但目前尚未观察到居民投放行为出现统计学意义上的显著改变；其主要制约因素在于各城市议会与管理机构仍在审议数据隐私政策与行政法规，面向公众的自动化反馈与惩戒机制尚未在市政层面全面推行。

Detailed notes — zh

## Background 该内容由频道 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台上。页面简介指出，本演讲由 Battle Motors 的首席软件官 Paul Marsolan 主讲，深入探讨了部署于 80,000 磅重型垃圾车上的 Fortris™ 边缘视觉 AI 平台，涵盖边缘多模型并行推理、低延迟决策、云端双向数据闭环以及在恶劣车载环境下的实战经验。 ## TL;DR Battle Motors 通过自研的 Fortris™ 软硬件一体化平台，将边缘 AI 视觉技术引入 80,000 磅的重型职业卡车中。系统在车载 NVIDIA Jetson 芯片上并行运行通用、专有及动作检测三个模型，实现低于 200 毫秒的超低延迟响应，精准识别锂电池等高危易燃物、垃圾污染及活体目标。结合双屏驾驶舱整合、SOP 级料斗火灾预警、智慧城市道路感知与人机协作（Human-in-the-Loop）云端数据闭环，该方案在无网或弱网环境下大幅提升了清运作业安全与运营效率。 ## Atmosphere & style 演讲整体呈现出极其专业、务实且充满工业实战气息的风格。演讲者语言诙谐幽默，用通俗易懂的语言将复杂的边缘计算架构与垃圾车日常作业的真实痛点巧妙结合。现场氛围沉浸且富有说服力，既有对重型卡车硬核制造与工业级安全的敬畏，又充满了将前沿 AI 技术落地于最严苛现实环境的笃定与自豪。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=4s) Battle Motors 位于俄亥俄州，是美国第二大职业卡车 OEM，专为垃圾清运等场景打造重达 80,000 磅的重型卡车。 - [00:41](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=41s) 传统卡车行业存在底盘厂、车身厂和后装市场拼凑导致的设备碎片化与驾驶舱多屏混乱问题，Fortris™ 旨在提供原厂一体化解决方案。 - [03:41](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=221s) 由于卡车常在缺乏 LTE 信号的环境下作业，Fortris™ 将 AI 视觉模型部署在端侧，实现低于 200 毫秒的本地推理与决策响应。 - [03:41](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=221s) 本地边缘计算能够实时识别丙烷罐、锂电池等高危易燃易爆品，并在机械臂倾倒垃圾时拦截，防止料斗起火。 - [03:41](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=221s) 系统通过实时监测和识别回收垃圾中的普通垃圾袋，帮助清运商防范违规污染并避免高额政府罚款。 - [05:58](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=358s) Fortris™ Hub 采用乘用车级硬件架构，配备 14.6 英寸中控屏与 10.2 英寸数字仪表盘，并通过 mTLS 双向证书加密保障车载与云端通信安全。 - [07:42](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=462s) 车上配置 8 个 RGB 摄像头与 NVIDIA Jetson 芯片，并行运行通用视觉模型、垃圾专有模型以及动作检测模型。 - [07:42](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=462s) 系统利用动作检测模型识别机械臂举升与倾倒动作，仅在特定作业瞬间激活高算力专有模型，有效降低功耗与存储占用。 - [09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) 面对火灾预警的误报风险，系统结合多帧特征校验与驾驶员人工确认（Human-in-the-Loop），并将确认数据自动回传用于迭代模型。 - [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 结合 GPS 和视觉图像，系统能够生成带影像证据的服务履约证明（Proof of Service），有效解决卡车未收垃圾等客诉纠纷。 - [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 作为每周遍布城市街道的移动感知节点，垃圾车可利用闲置算力自动检测路面坑洼、故障交通灯及受损路标，助力智慧城市管理。 - [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 卡车采集的图像数据通过 AI Pipeline 与 LLM 进行自动标注，打造高效率的模型迭代闭环。 - [21:22](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1282s) 在隐私保护方面，系统在数据管线中会自动对人脸和车牌进行模糊处理，且数据保留 14 天后即彻底删除。 - [24:00](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1440s) Fortris™ 提供灵活的商业化模式，支持工厂预装或后装，并提供硬件买断、软件 SaaS 订阅及 Hardware-as-a-SaaS 打包服务。 ## Detailed walkthrough 以下是针对演讲视频前半部分（第一部分）的详细且忠实还原的文字整理与技术剖析。 --- ### [00:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=4s) 垃圾车与边缘 AI 的交汇点：重型职业车辆的现代化变革演讲者以一个全场观众皆有共鸣的现象开场：每个人每天都会产生垃圾，也几乎每天或每周都会在自家门口见到垃圾车。垃圾被扔进车后仿佛就“凭空消失”了，这构成了现代社会高效运转的基石。然而，在这看似平凡的日常背后，蕴含着极大的科技升级潜力。作为美国第二大职业卡车原始设备制造商（OEM），Battle Motors 拥有位于俄亥俄州新费拉德尔菲亚（New Philadelphia）、占地约 32.5 万平方英尺的制造工厂，专为垃圾清运等场景打造重型卡车。作为软件首席官，演讲者正带领团队通过 AI 视觉技术重构这重达 80,000 磅（约 36 吨）的重型机器，旨在显著提升驾驶安全、优化运输路线、保护行人并防止垃圾污染。传统的垃圾车设备极其碎片化：车厂先制造出底盘驾驶室（Cab & Chassis），再将其交付给车身制造厂（Body Maker）加装后部的金属垃圾箱，最后运抵洛杉矶等城市。各城市或运业商随后自行加装各种后装市场（Aftermarket）的摄像头、导航软件和监控屏幕。这种层层拼凑导致驾驶舱内乱成一团，司机需要同时盯着多个独立屏幕。Battle Motors 的核心目标是提供一套高度集成的原生软硬件平台——Fortris™，将现代乘用车的智能体验（如倒车 360 度全景环视、盲区预警等）直接带入 80,000 磅的重型卡车中，并通过车载 GPU 运行本地 AI 模型，实时解决火灾防范、回收物污染检测以及行人碰撞风险等严重隐患。 - **工厂规模与定位**：位于俄亥俄州新费拉德尔菲亚，厂房面积 325,000 平方英尺，为美国第二大职业卡车 OEM。 - **车辆基本参数**：典型总质量（GVWR）达 80,000 磅（约 36 吨）的重型职业卡车。 - **行业痛点与现状**：传统的“底盘车厂 → 车身改装厂 → 客户后装”流程导致技术架构高度碎片化，驾驶舱内缺乏统一交互界面。 ``` [底盘制造 (Battle Motors)] ──> [车身加装 (Body Maker)] ──> [后装加装 (第三方系统)] │ (碎片化、多屏幕) ▼ [Fortris™ 一体化原生架构] (360环视 / 车载GPU / 单屏) ``` --- ### [00:41](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=41s) 离线边缘计算与本地 sub-200ms 高实时响应在重型卡车的实际作业中，通信环境极其恶劣。垃圾车穿梭于城市的各种复杂角落或偏远地区，经常遭遇 LTE 蜂窝网络信号中断或完全无网的情况。如果依赖云端处理，一旦网络断连，关键的安全预警就会失效。因此，Battle Motors 选择将所有核心视觉模型直接部署在卡车本地。这套车载边缘计算系统能够在无网络连接的状态下独立运行，实现低于 200 毫秒（sub-200 ms）的模型推理与决策响应。这种超低延迟的本地计算能力，对于实时拦截高危物品至关重要。例如，当垃圾车机械臂将垃圾桶倾倒进车厢时，系统必须在百毫秒内识别出桶内混入的丙烷罐（液化气瓶）、锂电池或笔记本电脑等易爆易燃危险品，并及时发出警报甚至联动控制机械臂停止压实（Compact）操作，避免造成爆炸或毁车事故。此外，在加利福尼亚州等地区，法律对回收垃圾和堆肥的污染率（Contamination Rate）有严格规定（通常允许 20%–30% 的误投容忍度），一旦超出标准，清运商（Hauler）将面临高额罚款。通过边缘 AI 实时识别混入回收桶的普通垃圾袋，清运商不仅能避免罚款，还能向违规投放的住户提供精准教育乃至追溯罚款。 - **边缘推理延迟**：本地 AI 模型推理与决策响应时间完全控制在 **200 毫秒以内**。 - **关键高危识别目标**：丙烷罐、汽车铅酸电池、锂电池、笔记本电脑、手机。 - **合规与罚款防范**：针对加州等地的法规，防范回收物与堆肥中超过 20%-30% 的垃圾污染，保护清运商免受政府罚款。 --- ### [05:58](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=358s) 乘用车级硬件架构：Fortris™ Hub 与双屏安全网络为了彻底改变传统卡车满屏悬挂的乱象，Battle Motors 研发了全汽车级（Automotive-grade）、自研从零打造的硬件载体——Fortris™ Hub。这套硬件的规格向 Tesla Model X 等高端乘用车看齐：中控台配备一块 14.6 英寸的大型触控屏，驾驶员前方则设有一块 10.2 英寸的数字仪表盘（Instrument Cluster）。通过这两块屏幕，司机可以直观地获取 360 度环视图像、盲区报警以及当前的作业上下文信息，从而集中注意力于驾驶本身。在数据安全方面，Fortris™ Hub 与云端“命令服务器”（Command Server）之间的通信采用了双向证书认证的双向 TLS（mTLS, Mutual TLS）加密协议。演讲者幽默地强调，这套卡车系统的网络安全级别“甚至比你手机登录银行或者银行自身的系统还要安全”。通过 Factory-installed（出厂预装）的闭环体验，车队运营后台可以随时在网页端查看每辆车的实时状态、行驶路线，甚至能追溯特定住户（例如“雅各布先生家”）的垃圾收集记录，瞬间厘清住户是否向普通垃圾桶里偷扔了汽车蓄电池。 - **显示硬件规格**：14.6 英寸中央控制屏 + 10.2 英寸驾驶员数字仪表盘。 - **网络加密协议**：车载 Hub 与云端命令服务器之间采用 **mTLS（双向 TLS）** 建立高安全性通信隧道。 - **多屏融合体验**：整合倒车影像、360 度环视、盲区监测（Blind Spot Indicator）与实时路线上下文。 --- ### [07:42](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=462s) 多模型并行推理与基于机械臂动作触发的视觉过滤机制 Battle Motors 的垃圾车外部配置了 8 个 RGB 摄像头，所有视觉数据流均汇聚至车载 NVIDIA Jetson 芯片（Nvidia Orin 系列）。在芯片上限的算力约束下，系统并行运行着三个不同分工的 AI 视觉模型： 1. **通用模型（Generalist Model）**：负责开放世界（Open-world）的环境感知，高效识别行人、猫、狗等目标。演讲者特别提到，在纽约等寒冷城市，无家可归者常会爬进大型垃圾集装箱（Dumpster）中避寒取暖，清晨垃圾车倾倒垃圾时极易发生将人误倒入车内并压实的惨剧。通用模型能及时识别活体，防止悲剧发生。 2. **专有视觉模型（Specialist Model）**：专门针对垃圾清运场景训练，精确分类各种废弃物（如塑料袋、玻璃、金属、丙烷罐、电池等）。 3. **动作检测模型（Action-detection / Arm Model）**：用于监测机械臂与车厢的物理动作，实现智能事件过滤。由于卡车全天候行驶会产生海量视频数据，如果全程持续进行复杂的 AI 深度分析，会造成极大的算力与存储浪费。因此，系统引入了类似家用 DVR 的全天候背景录像，但对于高算力的 AI 视觉分析，则由“动作检测模型”进行精准控制：只有当自动侧加载机械臂（Automatic Side Loader）执行举升、倾倒垃圾的“ mission（任务）”瞬间，系统才激活专有模型去扫描集装箱/料斗（Hopper）内部的物品。这种“按需触发”的机制大大降低了芯片功耗并提高了检测精准度。 ``` [8路 RGB 摄像头] ──> [NVIDIA Jetson 车载芯片] │ ┌──────────────────┼──────────────────┐ ▼ ▼ ▼ [通用模型 (人员/宠物)] [动作检测模型 (机械臂)] [专有模型 (垃圾/高危品)] │ │ │ └─────────┬────────┘ │ ▼ ▼ (监测是否举升/倾倒) ──> (仅倾倒时激活) ──> [料斗内部高精度分析] ``` - **传感器配置**：8 个 RGB 摄像头（曾实验过热成像与单纯视觉，最终选定视觉摄像头组）。 - **车载计算芯片**：NVIDIA Jetson 平台（运行多模型并行推理）。 - **计算过滤策略**：全天候 DVR 录像 + 基于机械臂举升动作触发的 AI 模型推理，极大节省算力。 --- ### [09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) 车厢料斗火灾分级预警与人机协作（Human-in-the-Loop）数据闭环料斗内部火灾是垃圾清运业最头疼的灾难之一，通常由误投的锂电池被挤压破裂或未熄灭的灰烬引发。演讲者将车厢火灾归纳为三个渐进演化的严重等级： 1. **初起阶段（烟雾/微火）**：系统捕捉到早期烟雾或微小火苗，立刻给司机弹出警告提示，并按车队的标准 SOP（标准作业程序）给出操作清单。司机可找安全地点停靠、压实车厢，并观察 20–30 分钟等待火势自行熄灭。 2. **局部蔓延阶段（卸料救车）**：火势扩大可能吞噬整车，司机需要紧急寻觅空旷地点（如空旷停车场或死胡同 cul-de-sac）将整车着火的垃圾倾倒入地面。这种“在居民区倾倒一整车着火垃圾”的做法极其昂贵且极不受欢迎，但能保住车辆。 3. **资产全毁阶段（毁灭性火灾）**：整辆价值不菲的卡车被完全烧毁，车队面临巨大的财产损失和运力缺口。为了在第一阶段就将火灾消灭在萌芽状态，Battle Motors 利用真实料斗背景合成（Synthetic Images）了带烟火的数据集并训练出专有模型。然而，工地扬尘、橙色垃圾袋等极易引发误报（False Positive）。为了解决这一难题，系统采用了双重机制：首先要求烟火特征必须**持续多个帧（Sustained for multiple frames）**才触发预警；其次引入了“人机协作（Human-in-the-Loop）”反馈——当屏幕弹出火灾预警时，司机会确认“这确实是火灾”或“这只是扬尘误报”。司机的每一次手动确认或否定，都会自动打包上传至云端，直接转化为带标签的高质量训练数据，不断迭代优化本地模型的鲁棒性。 - **火灾演进三阶段**：初起烟雾（SOP指导） → 局部蔓延（紧急倾倒卸料） → 资产全毁（整车烧毁）。 - **误报消除策略**：多帧连续特征校验 + 驾驶员人工交互确认（Human-in-the-Loop）。 - **数据反馈闭环**：司机的纠错反馈直接回传至云端模型训练管线（Training Pipeline）。 --- ### [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 服务履约证明、超载检测与城市级“移动感官”节点除了安全与防火，Fortris™ 平台还极大地改善了清运商与末端住户之间的客诉矛盾。住户经常打电话向客服投诉“垃圾车今天没收我家的垃圾”，而司机则辩解“是他根本没把垃圾桶推出来”。为了化解这种无休止的扯皮，系统通过 GPS 与视觉系统结合，实时记录垃圾车经过住户门口的视频或照片证据（Proof of Service）。一旦发现住户没摆放垃圾桶，系统会自动生成带影像证据的链接或邮件发送给住户，类似于闯红灯摄像头的“铁证如山”，大幅降低了客服部门的处理成本。此外，系统还能通过 AI 视觉识别商业或民用垃圾桶的“超载（Overfilling）”现象，帮助清运商进行合规计费。将目光放得更远，每周都会驶过城市每一条街道的垃圾车，本质上是一个极其理想的“城市基础设施移动感知节点”。Battle Motors 正与合作伙伴共同探索“智慧城市”扩展应用：垃圾车在执行清运任务的同时，利用闲置算力和摄像头自动检测路面的坑洼（Pothole）、损坏或熄灭的交通信号灯、被毁坏的停止标志（Stop Sign）等。与其专门派市政车辆无目的巡逻，不如利用垃圾车这一天然的流动网格，将城市问题自动汇总至市政管理部门。演讲者随后展示了他们的数据训练管线架构：卡车采集到的原始图像回传后，通过 AI Pipeline 并引入大语言模型（LLM）进行自动标注（Auto-labeling）与目标框定（Bounding）。虽然演讲者笑称自己有时也得亲自进行无聊的人工标注，但现代 LLM 与计算机的自动化标注效率远超人工，标注好的数据集被迅速投喂给模型更新，再通过云端下发回每辆卡车，形成完整的智能化进化飞轮。 - **服务履约证明（Proof of Service）**：结合 turn-by-turn 导航与视觉图像，向未摆放垃圾桶的住户提供直观视频/照片证据。 - **智慧城市感知能力**：自动识别道路坑洼（Potholes）、故障交通信号灯及受损路标。 - **数据训练管线**：卡车原始图像 → AI Pipeline + LLM 自动标注/画框 → 模型训练与迭代。 ### [17:43](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1063s) 自动驾驶辅助与驾驶舱整合云端验证系统会将新训练的模型与车辆当前运行的模型进行对比，基于训练数据和验证测试结果评估是否更新，每辆车每天可处理并训练数以万计的图像。典型的垃圾车每天需要停靠约 1,000 个站点（覆盖 1,000 户家庭），产生了海量的场景数据。在视觉 AI 系统之外，车辆还集成了高级驾驶辅助系统（ADAS）等安全关键功能。自动紧急制动（AEB）等功能的引入，极大地提升了市政部门和运输承包商的作业安全。同时，结合驾驶员辅导和 DVR 事故保护功能，视觉数据得到了充分利用。演讲者通过对比传统垃圾车与配备 Fortris™ Hub 的车辆驾驶舱，展示了原厂制造（OEM）的软件整合优势： ``` +-----------------------------------------------------------------+ | 传统驾驶舱 (3-4 个屏幕) | | [ 路线导航 Tablet ] [ 多摄像头显示屏 ] [ 垃圾收集作业屏 ] | +-----------------------------------------------------------------+ VS +-----------------------------------------------------------------+ | Fortris™ Hub (仅 2 个屏幕) | | [ 上下文整合显示屏 1 ] [ 上下文整合显示屏 2 ] | +-----------------------------------------------------------------+ ``` - 传统驾驶舱内部通常充斥着 3 至 4 个独立屏幕：一个用于路线导航，一个用于查看多路摄像头，还有一个用于跟踪垃圾清运任务。 - Fortris™ Hub 将所有关键信息上下文整合至仅 2 个屏幕中，直接在出厂时完成预装，避免了后期零散拼凑的问题，使驾驶环境更加整洁安全。 --- ### [19:49](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1189s) 问答环节：误报率与摄像头安装位置在问答环节中，针对现场观众提出的模型倾斜度与误报率（False Positives）问题，演讲者表示在锂电池引发火灾的检测上，系统更偏向于容忍误报。因为垃圾车起火属于灾难性事件，虽然发生频率不高，但危害极大。因此，系统宁愿将疑似情况标记为火灾并提醒驾驶员确认，给操作员提供具体的现场上下文。关于摄像头的物理安装位置： - 摄像头的安装取决于具体的配置与应用场景。例如，用于客户清运验证的摄像头通常安装在车尾的散热器附近。 - 视具体捕获需求，摄像头也可能直接安装在车身上。 - 盲区监测摄像头通常安装在驾驶舱侧面，一般不会直接挂载在后视镜上。 --- ### [21:22](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1282s) 问答环节：隐私保护与边云协同权衡针对私人领地拍摄与隐私合规的提问，演讲者解释道，保护隐私是平台的核心职责。系统在数据管道传输过程中会自动对人脸和车牌进行模糊处理（类似于 Google Street Maps 的做法），并且所有数据在保留 14 天用于保障数据完整性后会被彻底删除。在回答为何选择端侧芯片（On-chip）而非大模型云端推理时，演讲者指出了两个核心限制因素： - **网络连通性：** 许多偏远或乡村地区的垃圾车缺乏 LTE 信号覆盖。如果车厢起火，不可能等到车辆驶回有 T-Mobile 信号的区域再发出警报。 - **成本与实用性：** 云端大模型的推理成本极高。端侧配备本地 GPU 可以有效控制成本，同时实现亚秒级的实时响应（如实时防火、预防性通知超载等）。对于非实时需求，云端推理仍可作为补充。现场有人询问居民在得知被监控后是否改变了丢垃圾的行为（如减少超载），演讲者指出目前尚未看到明显的统计数据变化。这主要是因为合作城市尚未全面推进居民科普与教育流程，市议会在批准相关隐私与告知条款时依然较为审慎。 --- ### [24:00](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=1440s) 商业化模式与定价策略 Fortris™ 平台采用了灵活的硬件与订阅制（SaaS）组合商业模式： - **硬件获取方式：** 客户既可以选择在购买新车时直接从工厂预装硬件，也可以通过后装（Aftermarket）方式升级现有车队。硬件支持一次性买断，也可按月分期。 - **软件与服务：** 平台的核心功能以订阅服务方式提供。 - **硬件即服务（Hardware-as-a-SaaS）：** 类似于手机合约套餐，客户可以将硬件成本与软件订阅整合为每辆车每月固定的打包服务费。 ## Notable quotes > [05:58](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=358s) 甚者比你手机登录银行或者银行自身的系统还要安全 — Paul Marsolan > [09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) 这确实是火灾 — Paul Marsolan > [09:55](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=595s) 这只是扬尘误报 — Paul Marsolan > [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 是他根本没把垃圾桶推出来 — Paul Marsolan > [13:04](https://www.youtube.com/watch?v=gdUAQLIMqxM&t=784s) 垃圾车今天没收我家的垃圾 — Paul Marsolan ## Mentioned Paul Marsolan, Battle Motors, Fortris™, Fortris™ Hub, Ai4, NVIDIA Jetson, NVIDIA Jetson Orin Nano, Tesla Model X, Google Street Maps, T-Mobile ## Video description Paul Marsolan - Chief Software Officer, Battle Motors What does it take to ship production AI on an 80,000-pound refuse truck — a vehicle that fights vibration, heat, dust, and intermittent LTE on every shift, and where a single missed detection can mean a hopper fire from a discarded lithium battery? This talk goes inside Fortris™, Battle Motors' edge vision platform: eight RGB cameras, three vision models running in parallel on a Jetson Orin Nano, sub-200 ms decision latency, and a bidirectional cloud plane that turns every route into a labeled dataset. We'll walk through the real architecture — specialist + generalist + action-detection models, OTA model deployment with per-truck rollback, ADAS sensor fusion — and the production lessons behind it. Deploying a model to 1,000 trucks is not deploying to 1,000 web servers; reversibility is a product feature, not a nice-to-have. You'll leave with a concrete picture of what edge AI looks like when it has to work in one of the harshest commercial environments in America — and why the cab of an 80,000-pound truck might be the next great human-AI interface. [Source](https://www.youtube.com/watch?v=gdUAQLIMqxM)

Source quotes

[5:58] Paul Marsolan: 甚者比你手机登录银行或者银行自身的系统还要安全 [9:55] Paul Marsolan: 这确实是火灾 [9:55] Paul Marsolan: 这只是扬尘误报 [13:04] Paul Marsolan: 是他根本没把垃圾桶推出来 [13:04] Paul Marsolan: 垃圾车今天没收我家的垃圾

Source quotes

[22:10] Paul Marsolan: I can't be like, 'Hey, there's a fire in your truck, but only when you get back to T-Mobile cell phone service.' That doesn't really work out very well.

Original transcript

[0:04] So, everybody in this room has something in common. We all make trash and we all see a garbage [0:09] truck almost every single day or at least once a week and it shows up to your house and you pick up [0:15] stuff and you throw it in the trash and you never see it again. And that makes it for a wonderful, [0:20] wonderful world. Now, the really cool part about this is how can AI help driver safety [0:26] and haulers to to make things more optimized and also safer for not only the environment but [0:35] also for the trucks and us as pedestrians. So, like we talked about before, Battle Motors. So, [0:41] Battle Motors is the second largest OEM vocational truck manufacturer here in the United States. So, [0:47] we have a factory in New Philadelphia, Ohio. About 325,000 square feet there. And we make [0:54] vocational trucks. Uh, this is one of our our garbage trucks you see right here. And I am in [1:01] charge of all the software there. And so one of the cool problems that I have to solve is [1:06] how do I prevent contamination for recycling and and other things inside of just regular [1:13] trash versus recycling versus compost along with how do I stop emergencies from happening like [1:20] fires? And the way today people see that is they usually get stopped at a red light and it's like, [1:25] "Hey bud, your truck's on fire." And you potentially will lose this entire vehicle. [1:29] But with some of the technology we've developed, we're able to use AI vision to be able to not [1:34] only help that use case, but also help with pedestrians and help with also contamination. [1:42] So [clears throat] [1:45] let's see what we have. So today we have eight cameras and we we've experimented [1:49] with thermal cameras. We experimented with vision cameras um and really found a lot [1:55] of success with the vision camera sets and a lot of vision cameras exist today [2:00] uh via an aftermarket solution. So the way the refuge market works is we build a cabin chassis, [2:05] you deliver this cabin chassis to a body maker. The body maker puts the thing that looks, [2:10] you know, like a box in the back of your truck and then it gets shipped to off to [2:14] wherever it's going to go to. So let's just say the city of Los Angeles. And in this case, [2:20] they then add in all these cameras and routing software and whatever else it may be. So, [2:24] it's very fragmented. My goal was to be able to create one consolidated platform to be able [2:31] to have it completely integrated into the truck. No longer numerous screens looking around to see [2:36] where the driver's going, what's he doing, what's he looking at, what cameras he looking at. I want [2:41] the same thing that's in our residential vehicles put in this massively large truck cuz why not, [2:45] right? We push reverse, we see a 360 view of our cars. Why not in an 80,000lb truck? So, what we're [2:53] able to do now is have all these cameras be able to feed into a centralized dash, which I'll show [2:58] you in a second, but then also add in AI on top of it locally with a GPU in the truck to [3:04] be able to solve a lot of different problems. And some of those problems exist with with fires like [3:10] we talked about earlier, contamination, people. You'd be surprised about how many garbage trucks [3:17] hit people yearly. Uh so being able to alert the drivers, be able to stop the truck appropriately [3:24] and then be able to use this information for coaching in the future and bring all this data not [3:29] only it it be locally in the truck itself for fast response times of under under 200 milliseconds, [3:36] but also be able to pass this data back up to the cloud for training and coaching. And so being able [3:42] to have this full circuit of what did the truck see, what is the exper environment around the [3:48] truck, and then what are we doing with the data to be able to identify these use cases and help [3:54] educate and coach drivers to become safer and mitigate the things they're trying to do today, [3:59] such as looking, did we get everything in the hopper? Did we pick up a battery? Do we pick up [4:03] a propane tank? being able to identify those with the vision system and allow the driver to do the [4:08] thing they're supposed to be doing, getting the truck from point A to point B safely. [4:14] So, one of the things that we tried to do here was garbage trucks go all over the place and there [4:20] might not be cell phone signals in all those locations. So, you have to do things locally. [4:26] Well, you don't have to, but that's the way we decided to do it. [laughter] So, do it locally for [4:31] speed. So again, our models run in less than 200 milliseconds locally on the truck itself. And then [4:36] we're able to not require cell phone connectivity. And so what that means is we're able to identify [4:41] fires. We're able to identify dangerous items. You picked up a propane tank. Don't compact. You don't [4:46] want to explode the truck. Identify different things like laptops. Identify contamination. [4:51] In California, we have a regulation that haulers get fined for if there's too much contamination [4:59] in like a recycling bin or a compost bin. So if you're putting your trash bag in a recycle bin, [5:04] the hauler physically can get fined for that because they're dumping this contamination [5:08] into this recycling route. And so being able to identify this information and of course there's [5:14] right there standards that say there's 30% or 20% can be contaminated. That's perfectly fine. [5:20] But once you start to get above that, the hauler actually gets fined more and more. And so what [5:24] we're going to start to see, I think, and and some places already do this, especially with [5:27] private companies, is they actually will come back to you and say, "Mike, you've been putting [5:32] trash in your recycling bin. We're going to start finding you now for different things." So some [5:37] education and then some finding will actually take place. Um, and then all this data, again, [5:44] like we were talking about before, gets pushed up to the cloud for us to be able to use for training [5:48] and coaching and training our own AI models. So, how do how do we really do this? So, right now, [5:58] this is Fortress. This is Fortress Hub. We can put it in our this is what comes in our trucks. So, [6:04] a person who tries to buy our truck, think about it like a Tesla Model X. Got a 14.6 in screen in [6:09] the center and it's got a 10.2 in screen up front for the driver and operator in instrument cluster. [6:15] We can do a tremendous amount of cool stuff with this thing. So, it's all automotive grade. It's [6:20] all our hardware. We designed it from ground up. Uh and then all this reports back up to our [6:25] our command server is what we call it. And we're using mutual TLS in between it. So the nice part [6:31] is it's more secure than your bank is or more secure than your phone going to your bank. And [6:36] so we're able to use the cyber security on top of the system that we're able to install at the [6:41] factory to be able to bring in this cohesive experience for the operator. So it looks just [6:46] like our residential vehicles. And to put it in cont to put the contrast out there of what it [6:51] actually looks like today, most vocational tracks are all analog. This is the first of its kind to [6:59] really bring in the residential capabilities of 360 camera, reverse camera when you do it, [7:05] blind spot indicator, and bring this contextual information to the operator so they're focused [7:10] on the thing they're doing. And then at the same time, fleet operations is able to look [7:14] online at the web page and be able to know what they're doing at whatever time they're doing it. [7:18] Where are they in the route? What did they pick up at this route? Mr. Jacob's house. What did we [7:24] pick up there? Oh, he picked up a car battery. Why is that in the trash? So, things like that we're [7:29] able to identify and be able to share not only with the fleet operations, but then also back [7:33] to the operator in real time to be able to give contextual information about what's going on. [7:42] So, this is a picture of of one of our trucks and a whole bunch of the cameras that sit on top of [7:46] it. Now, the really cool part about this is we're running a couple different models simultaneously [7:51] on this on this Nvidia chip. And one model is is a journalist model. So, I'm looking for people [7:58] and and and dogs and cats and whatever the case may be. And you'll be surprised about how often [8:02] all of those things are in trash cans. Uh usually alive, like especially like in New York City where [8:09] it gets cold, people like to go to sleep inside of a dumpster if they don't have someplace to stay. [8:12] And guess what comes early in the morning? Refuge trucks. And so what ends up happening is they pick [8:17] him up and some poor guy's got a knock on the wall of the of the truck telling them he's he's been [8:21] stuck in there. Uh so you got to let him out and ideally don't compact them. Then we have another [8:26] one that's specialized for refu specifically. And so we're looking for propane tanks, we're looking [8:30] for car batteries, we're looking for laptops, we're looking for cell phones, uh we're looking [8:35] for any type of refu. So plastic bags, containers, glass, metal. And then we have a a way to filter [8:44] out all these images that we're taking and only focus on the piece that we care about, which is [8:49] when the arm is operating. So when did you pick up something? I don't necessarily care you stopped a [8:54] stoplight. I don't necessarily care that you did this other thing. We record all this information [8:59] just like a, you know, your video system at your house. We have a DVR of every single event that [9:03] takes place so you can go back and look at it for accidents. But for the vision AI model system set, [9:08] we only really care about what you did at that point of operation, right? The the mission for [9:12] lack of a better phrase. So the mission, the ARM picks up the the the container, [9:17] dumps it in or on automatic side loaders, and we have a model that looks at that arm and makes a [9:22] decision. Did it did it lift or not lift? And then based upon that, we can look inside the hopper. We [9:27] can see what's going on inside the hopper and determine that Mike put garbage bags in his [9:31] recycling. And shame on him. And then we can send help the city or the private company or whoever is [9:37] hauling with the hauler we can help them provide education to these end users and say hey we don't [9:43] sort it at the facility so please stop doing that. Uh yeah so that's part of our camera system. Now [9:55] we decided to do this all locally in the truck and the main reason is for this. So there's really [10:00] three main cases of fires inside of of trucks. One is a fire will start, you start to see smoke, [10:06] you start to see the actual flame. We can alert, we can identify the operator, let the operator [10:11] know that in fact there is a fire happening. And so what are you supposed to do now? And we can [10:17] actually give them the checklist of whatever their fleet recommends for operational in that use case. [10:22] So in most cases there's a you'll pull over, you find a place to stop, you find a place you [10:26] can dump out the load if required and you compact and you wait there for about 20 30 minutes hoping [10:32] the fire goes out. The next excellation is that it actually catches the whole truck on fire and [10:38] so you can dump out the truck and usually you try to do this inside of a empty parking lot. [10:42] Sometimes you don't have a parking lot to go to so you you would do this at a cult. You see this [10:47] in a lot of cases, but this is extremely expensive proposition, right? Dumping a truckload of garbage [10:52] that's on fire in someone's neighborhood. It's usually frowned upon. And then the last one, [10:57] which is obviously the worst case, is the truck catches on fire and you lose the entire asset. [11:02] And so the whole truck is gone and you got to get a new truck ASAP. And so we what we try to do here [11:09] is we're alerting the operator of what's going on before that happens. And sooner the better in this [11:16] use case, right? You see smoke, let the operator know. And so you see here, this is this is one [11:20] of our synthetic images that we've taken from a real hopper. I've added I've added some fire into [11:26] it for from a synthetic perspective and then had our uh our model look at it and say, "Hey, there's [11:32] some fire. There's some smoke." And then a really nice part here now is since this is so dangerous [11:39] for the operators and for the trucks and for the fleet that we like the operator to give input. Is [11:45] this actually a fire? Did I give you a false alarm because it was an orange bag? Did I just pick up a [11:50] construction dump site? So there was like a whole bunch of dust and so we thought it was smoke. So [11:55] we built things into the model to be able to kind of filter some of these things out. So if it it's [12:00] sustained more than a couple frames for example in this use case to mitigate false alarms but [12:05] there is going to be a use case that exists that we ask the driver hey does this is this actually [12:11] [clears throat] a fire and we'll alert the driver and tell them and then they can tell us and the [12:15] nice part about it is once they tell us it becomes more training data for us. Now, using using the [12:22] operator's content to be able to provide us data, we're able to train the model better to mitigate [12:28] these false alarms. And we're we're taking all this data again back up to the cloud and and [12:32] putting it through our through our models. And in in this use case, right, this is this is our [12:37] specialty model that's identifying fire and smoke. Uh the generalist fire or the generalist model [12:43] really again is mostly for the world. It's just open kind of an open world uh model that we're [12:48] using that runs very efficiently on the processor and then gives us a lot about a bunch of context. [12:56] So [clears throat] one of the other big problems that we have as operators is there we go. Tada. [13:04] Is how do you know you picked up someone's house? And the real question then becomes, [13:09] can I help with technology to stop this fight between the end consumer calling headquarters [13:17] and fleet and basically saying, "Hey, you didn't pick up my house." And the and the driver saying, [13:21] "Well, there was no garbage can out. I can't pick up this house." And so you have a fight now, [13:25] right, between the the driver and the customer. And it's never like a good never good win. So, [13:31] what we provide is this camera footage of you picking up the house or we're driving [13:36] past the house and no c no can out and so we can actually forward that to you via email or [13:42] or a link. And ideally, we're trying to get in front of you calling and saying, "Hey, [13:46] you didn't pick up my garbage." So, it's just like the red light camera. I don't know if anybody's [13:50] ever had the fun experience of getting a red light camera uh footage, but I certainly have. And um [13:57] my wife was nice enough to answer the or look at the look at the mail and be like, "Shame on you, [14:01] Paul. Look at this. [laughter] You really ran that red light." So, it's the same type thing here. [14:06] We're able to show this video proof and pictures to the consumer to be able to basically mitigate [14:12] this customer service experience that the hauler has to then handle. And then on top of that, now [14:18] we're looking at different cases of overfilling. And so some haulers like to over like to charge [14:23] for overfilling the container. So if it's too full, they'll they'll charge you more versus not. [14:27] And this is usually inside the commercial space, but it can happen inside residential space also. [14:32] In fact, we're doing a city right now that's very interested in knowing if they're [snorts] their [14:37] users are overfilling their garbage cans. And so they're they're we're training the [14:41] model right now on consumer grade bins to be able to determine if it's overfilled or not. [14:56] So again, this is proof of service again. So we're talking about the numerous different use cases [15:00] that take place here to be able to make sure that in fact everybody's doing what they're supposed [15:04] to be doing and capturing and detecting everything that's happening from this point. And then we have [15:10] a routing system inside of Fortress also that allows for us to do turn byturn directions to [15:15] different locations. And using that information, building on top of the AI model for for vision, [15:21] we're able to not only know where the truck is, where it's going to be going, whose house it's [15:25] going to be stopping in front of, and then using all the cameras around with AI vision to be able [15:30] to know what the environment looks like in that world, and then be able to take the images and [15:35] use that for proof. and then again be able to take this information and share this up to the cloud to [15:40] be able to work on our training models and improve the efficacy of of the models locally itself. [15:50] So what happens now becomes really really cool and powerful is we have all this all [snorts] these [15:56] cameras driving around a truck that drives around almost every single street in the city [16:01] every single week, right? What kind of data can we get from this to be able to help other [16:06] municipal operations inside that city? And we've had numerous conversations upon, hey, [16:12] there's a pothole. Can the truck identify there's a pothole? Oh, can the truck identify if the [16:16] red light is out? Can the truck identify if the traffic signal is not working? Can the tra can the [16:21] truck identify the stop sign's been taken down? So, being able to use some of these data systems, [16:26] we're we're working with uh partners to be able to figure out cool and unique ways to capture [16:32] data to be able to help optimize the city in a smart way because why have somebody just drive [16:39] around randomly throughout the streets when you have this truck that's driving around [16:42] every single week and almost hitting every single street? Now, that comes up with a [16:47] whole bunch of unique and interesting problems obviously as you know who wants to go fill in [16:51] the the pothole if the truck identifies a pothole. Uh so, but anyway, the cool part [16:56] is that we're able to have these conversations and use this information to help create a more [17:00] optimized city all from this platform that we're really getting a lot of value from. [17:11] So the next this this is really what our training set kind of looks like, right? So we're we're [17:18] able to take in this data set, raw images from from the truck itself. We pump it through the [17:24] AI pipeline. Uh we we put an LLM on top of it to be able to identify things inside the images and [17:30] and box it. So humans aren't doing a lot of that. I do a fair amount of labeling myself, [17:35] which is no fun, but computers are much better at labeling than I am. So computer can do most [17:39] of that work for us and then use that to be able to do the training of that information. And then [17:43] we compare it against the model that currently exists on the truck itself and determine if it [17:48] should be better or not based upon training data and validation tests. And we use this [17:53] vision pipeline to train thousands and thousands and thousands of images per truck per day. Right? [17:59] Most most refuge trucks have a thousand stops a day. Right? That's a thousand houses a day. [18:04] So we're looking at all those type all those cases per truck. Now one of the really cool things about [18:13] us is we have all these vision systems with AI on top of it and we have ADAS features. And so ADAS [18:20] is safety critical information just like we have in our cars. You're about to crash into somebody, [18:24] it might beep at you or you might stop. It might stop your car physically what's going [18:29] on depending what your car is. Bringing these things on automatic braking capabilities into [18:34] the vehicle itself has really really been able to help a lot of municipalities and haulers into [18:40] this truck and bringing in now not on not only on top of this this vision system on top for driver [18:46] coaching and DVR capabilities for protection of of haulers during accidents. We're able to [18:51] use again all this data to be able to bring this along. And then I want to show you a picture of a [18:57] cab on the right. You see a cab that looks like it's got like four maybe three or four screens [19:03] inside of it. This is what a typical refuge truck looks like today. So you got one screen, [19:10] one tablet that shows you your routing. You got one screen that shows you numerous different [19:14] cameras. You got another screen that shows what are you supposed to be doing for pickups versus [19:18] our right, which is our Fortress Hub. And you see there's only two screens. And those two screens, [19:24] you can put all your information right there for the context context of that operator. [19:28] And that's one of the really cool things about being the OEM and supplying the software is that [19:33] we're able to bring in all these cool technologies into the truck off the factory line, not piece [19:39] meal on together, and then using this to help make the world better, safer, and and wonderful. [19:49] All right, that's all I got. Any questions? I'm sorry. Repeat that. One of your slides, [19:54] you mentioned that you're more biased towards false alarms. Yeah. So, what's your acceptable [19:59] false, you know, false positives? Yeah, that that's a that's a good question. So, [20:06] for fires, we we lean towards this false positive perspective because I prefer the [20:12] operator to tell us if that happens or not, right? And a fire is usually a relatively [20:17] catastrophic event for these vehicles. Uh the nice part is they don't happen that often. [20:22] They do happen often enough that we that everybody cares about it, but not that not super often. So, [20:29] we we lean towards the fire if it took place, but we also let them have some context around that [20:35] also. Um, so I don't have an exact number for you, but we do lean towards this is potentially a fire. [20:52] Are the cameras on the front of the truck? Are they mounted on the mirrors or on the [20:55] A-pillar? Yeah. Uh, so it depends on the truck configuration and what you're trying [21:02] to solve for. So if we're doing customer verification, for example, we have mounted [21:06] them on the radiator behind the truck uh on the battle trucks. If it's a different truck, [21:11] we might mount it on the body itself depending what we're trying to capture in that use case. Uh, [21:16] but usually not on the mirrors. It's usually uh for the blind spot uh it's usually on the side of [21:22] the cab. Another question. So your customers are the owner operators of these vehicles, [21:27] right? Yes. How do you get around uh private property? Right. So do they have to convey [21:34] to their customers that you're running cameras on your trucks? How does that work? Yeah. Um, [21:41] that is a very good question and the answer is the haulers do most their stuff and our job is really [21:48] to provide privacy. So just like Google Street Maps kind of blurs out people's faces and things [21:52] like that and license plates, we're doing the same thing as it comes across the data pipeline [21:56] and then we disregard or we delete images uh after around 14 days. And so the goal really for us is [22:02] is kind of data integrity for that 14-day period and then we just kind of throw it away. Yes. I'm [22:10] sure there's probably lots of reasons why you would um go with a like onchip model. Um do [22:16] you think that you could get similar results with a cloud based model like a bigger one and [22:21] I guess what are the trade-offs there? Yeah, so that's that's a good question and in fact that's [22:26] where we initially went. We went to we wanted we wanted to do a cloud inference case. Uh the [22:32] question really became cost and and then how much more value do you get from that cost perspective [22:38] from an LLM? And then the real the real kicker really was we have a lot of municipalities or [22:43] or customers that kind of live in rural areas and they don't have cell phone signal and so I can't [22:49] be like hey there's a fire in your truck but only when you get back to T-Mobile cell phone service [22:54] uh that doesn't really work out very well. So uh but things that are nonreal time we could [22:59] absolutely use cloud inferences if we wanted to. uh really just becomes a cost and so we try to [23:04] mitigate the cost perspective by just having the GPU local um by notifying the customers hey you [23:09] guys are overfilling right the proactive send out of hey we couldn't pick up today because [23:14] of these issues rather than them calling and say hey you didn't pick up my trash right yeah yeah [23:19] uh we [sighs] that's a that's a good question the answer is no and I have not seen any statistical [23:25] change in that particular uh use case and that's mainly because the cities that we're working with [23:30] have not moved forward forward with the education process yet. And so that's that's one of the [23:35] things that they're working past city council and this goes back to the privacy question that [23:39] was asked prior was like do we want to tell them this yet? And everybody wants to in the [23:45] in the refugees world. The question is the city council actually want to approve it and so yeah [24:00] what's your strategy in terms of pricing this? Are you doing it [24:03] subscriptionbased? Is it a one-time cost? Yeah, good question. So it's subscription based. So, [24:09] uh you can buy the hardware right from us from the factory installed in the truck or you can [24:14] get it on aftermarket. Either use case is fine. Uh and then we have a subscription service for [24:20] for all the cool stuff you see here. And then you can also either buy the hardware outright [24:25] much like your cell phone plan. You can either buy it outright or you can have it as a SAS [24:29] hardware as a SAS product set. So, it kind of lumps it into a monthly service fee per truck.