IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on Englewood’s cash-out + Axon 3–3, Marinette’s 9–0 after a bag, Lansing moving from cancel Flock → ban ALPR, and a real Jetson Thor box with 8× GMSL2.
Notable trends
- Kill-path is sharing answers Flock wouldn’t give (Englewood) or cover then 9–0 (Marinette).
- Keep-path is supervisor-only + 7-day + no ICE (Milwaukee).
- Policy is now ban the category (Lansing), not just the vendor.
- Edge news is a Thor cobot/NVR box, not Orin Nano 2 reprints.
- Lab is procedure VAD and real-time VLA editors.
Gaps
No new Hailo / DeepX / Qualcomm camera-NPU SKU. Orin Nano 2 is an Aug 25 recirculation. No new Genetec/Milestone SKU (Robox is an overlay). Half Moon Bay page was paywalled.
Mix counts
LPR 3 / VMS 1 / Edge 1 / Policy 1 / Lab 2 (LPR+Policy = 4/8 = 50%).
Now vs On deck
Now 6 (items 1–6) · On deck 2 (PROVIA, Real-Time EXPO-FT).
Signals (8)
- 1
Key point
Council unanimous terminate. Mayor Othoniel Sierra: talks since Oct 2025; staff never got a clean answer on how other jurisdictions use Englewood pings (they could pull more than the plate that hit). No local-officer misuse. City attorney: $27,500 to break the contract + $6,000 for Flock to pull the poles. As of airtime the stationary readers were still up; Flock did not answer Denver7 on a removal date. Next item: Axon ALPRs on Evidence.com (city-controlled servers) — 3–3, so paused. Chief David Jackson wanted some fixed ALPR kept: auto-theft 600+ (two years pre-Flock) → 507 (2023 launch) → 448 (2024) → 224 (2025). In-car ALPRs stay. Sierra: hope to talk policy (siting, neighborhood setbacks) this year or early next.
Why it matters
Kill-path is sharing, not local abuse. Quote the $33.5k cash-out and Axon 3–3 as the replacement vote that failed. Bid in-car vs pole as two SKUs.
Source
Denver7 - 2
Key point
Council 9–0 to end the contract. Clerk confirmed the count. Comes weeks after the mayor and police chief ordered the cameras covered while they reviewed citizen concerns. Same Fox Valley week as Green Bay’s 7-day keep-policy and De Pere’s bag-to-expire (numbered 09-15).
Why it matters
Cover-first, vote-second is now a WI playbook. Don’t wait for a misuse finding to write the bag clause.
Source
FOX 11 - 3
Key point
Public Safety Committee heard ACLU-MI + advocates pointing at Ypsilanti’s 2022 municipal ban. Goal: close loopholes so private vendors or outside agencies cannot keep pulling travel data. Councilmember Peter Spadafore (who voted against the Aug 24 7–1 Flock-cancel resolution on process grounds) now says council is examining a ban ordinance; charter-aligned draft 6–8 weeks. Aug vote already asked Mayor Andy Schor / LPD to terminate 26 cams at 7 sites and sign no new AI-ALPR until public engagement + privacy impact + policies. This is the next step: ban the category, not just Flock.
Why it matters
Michigan keep/kill is becoming ordinance-not-vendor. If you bid ALPR in Lansing, assume a ban window, not an Axon swap.
Source
WILX - 4
Key point
Fire and Police Commission reviewed MPD’s new guardrails after three officers investigated for misuse (including an IA officer on the first case). Updates: access supervisors only (no officers/detectives); cut outside-agency share; 7-day retain; ban civil-immigration searches. Chair Miriam Horwitz asked if data had ever gone to ICE; MPD did not give a clean yes/no on-air. Chief of staff Heather Hecimovich Hough: “Would we like to keep Flock? Yes… if we get a new vendor, that means starting that process all over.” Not a kill vote. Common Council still has an Sept 30 Q&A deadline.
Why it matters
Same state as Marinette’s 9–0 kill: big-city keep + shrink the graph. The close is supervisor-only + 7-day + no ICE, not a new camera.
Source
FOX6 - 5
Key point
Robox AI Video Intelligence Platform: on-site edge appliance on compatible IP streams + central software for config, alerts, evidence, multi-site analytics. No camera rip. CEO David Solomon: platform should evolve (refine models, add detections) not sit still. POS/transaction join for sweethearting, voids, refunds. Verticals: retail/grocery (LPR listed), QSR, warehouse/DC (PPE, forklift, worker-down), schools (weapons, fights — “subject to policy/law”). Eval → POV → commercial now.
Why it matters
If the customer will not replace Genetec/Milestone this year, the SKU is an edge overlay + POS join. Ask which two detections they will actually fund.
Source
SourceSecurity - 6
Key point
DeviceEdge AIE-KT78: NVIDIA Jetson T5000, 128 GB 256-bit LPDDR5X, up to 2,070 FP4 TFLOPS. AIE-KT68: T4000, 64 GB, up to 1,200 FP4 TFLOPS. Local LLM / VLM / VLA. Up to 8 GMSL2 cameras, QSFP28 (to 4×25 Gbps), dual 10 GbE, EtherCAT to motors/joints. Aimed at cobots, humanoids, AMRs that have to perceive and act on one box.
Why it matters
This is a Thor NVR/robot brain, not a locked AI dome. Spec GMSL2 + EtherCAT so you can swap the VLA later. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
Source
RoboticsTomorrow; Aetina - 7
Key point
Assistants should flag a mistake from past frames only, before the next step, and keep working after recovery. Current “first-mistake” benches cut the video at the first error — a timer that never looks at pixels wins. PROVIA keeps a factual state (what was actually done, mistakes included) apart from accepted progress (automaton posterior from correct demos). Sequential test → alarms under a false-alarm budget. Best among controlled baselines on CaptainCook4D, IndustReal, HoloAssist, IMPACT-ego. At 0.1 FA/min: recall .154 vs .128 (CaptainCook4D), .034 vs .015 (HoloAssist). 58–70 fps. Code: github.com/Kratos-Wen/PROVIA.
Why it matters
Plant/SOP cameras need procedure VAD, not another locked “AI camera” that only fires on a person in frame. Spec online state, not a clip that ends at the first error.
- 8
arXiv: Real-Time EXPO-FT — VLA proposes, a fast editor reacts; 42% → 97% in 10 minutes of robot data
Key point
Slow pretrained VLA emits action chunks; a lightweight edit policy patches them from the latest observation. Instantiation of EXPO-FT for real-time. Kinetix: delayed policy best in 10/10 envs vs delayed and non-delayed. Four real-world tasks (object pass, ball balance, table-soccer kick, dynamic pick): 10 minutes of online robot data, no human intervention, average success 42% → 97%.
Why it matters
Cobot/AMR RFPs fail on reaction time, not on a demo take. Keep the big VLA, add a fast editor on the Thor box — don’t buy a new locked camera brain.