IVS / Edge AI Research Brief

Signals (8)

1. 1

   Boston abandons Flock — sharing toggles unset within 3 days of a no-cost pilot

   Key point

   BPD’s 2025 Surveillance Technology Report (due to council Monday) says the 156-day no-cost Flock pilot (Apr 1–Sept 4 2025, 45 cameras) contracted sharing off; BPD confirmed settings off. Within three days other agencies could read Boston plates — Flock called it an “error.” Then 77 outside-agency requests (State, ATF, DEA, Brookline, Watertown, MSP, Maine SP, plus ~24 internal BPD asks). Officers ran 11,004 searches; ~half had no case number. No civil-rights complaint finding; data deleted and hardware removed within 30 days of trial end. Mayor Michelle Wu: the city has “abandoned Flock”; will not use a platform whose toggles “can be unset.” Not an ALPR exit: later Motorola (45 cams, Feb 2026) and Axon (30 cams, June 2026) pilots. ACLU-MA: ≥25 MA munis off Flock in 18 months.

   Why it matters

   The kill is vendor cannot keep sharing off, not “we don’t want plates.” Quote Wu’s toggle line in every New England RFP. Map the replacement stack: Motorola + Axon already in the same city.

   Source

   WBUR; GovTech / MassLive
2. 2

   Gadsden County, FL 4–1 — Flock dies Oct 1; no replacement named

   Key point

   Commission authorized written non-renewal of the sheriff’s Flock deal when the current term expires Oct 1, 2026. First deal March 2024. County has paid $162,000; ending it saves >$77,000 every two years. Sheriff Morris Young had already suspended deputy access. Staff: a motion was required or the agreement auto-continues. 4–1 (Commissioner Shawn Wood no): Wood, a former LE bureau chief, said cameras helped 3–4 homicides and warned “when little 4-year-old Susie gets kidnapped.” Commissioner Ronterious Green tied the vote to the sheriff’s rec and DeSantis’s order to pull Flock off state-maintained highways. No substitute vendor named; staff will look.

   Why it matters

   North Florida is now state ROW + county PO both going dark. Bid the replacement camera / grant question the board asked out loud — they killed Flock without an ALPR successor.

   Source

   WCTV; Fox 49
3. 3

   Marshall County, IN Council 4–3 defunds 6 county-paid Flock cams

   Key point

   Council 4–3 ended funding for six county-paid Flock cameras. Prosecutor’s $20,000 2027 request got $0. Warrantless-search concern; outside agencies ran hundreds of thousands of searches in one month. Contract ran through 2028; early exit $500/camera. Commissioner Jesse Bohannon: 25 cameras in the county; Bremen and Plymouth still have their own contracts — tonight only kills the county-paid poles. Takedown date for the six is unknown. Follows Kosciusko County.

   Why it matters

   Same county, three buyers. Quote county vs town contracts on the map. The $500/cam break fee is the real close number, not the 2028 end date.

   Source

   WNDU; ABC57
4. 4

   Green Bay keeps 42 Flock cams (retain 30 → 7) — De Pere bags until Nov 15

   Key point

   Same Brown County, opposite forks. Green Bay: Mayor Eric Genrich + Chief Chris Davis keep 42 cameras (in use since 2022); no misuse finding. Draft Policy 424: retain 30 days → 7 (except active cases); searches need a case number and incident call; share only with agencies that have similar policies; monthly public report + audit by the 15th. 30-day comment through Oct 14 — Davis: not a keep/kill referendum. De Pere: chief will bag immediately and let the $78,500 / 3-year contract expire Nov 15 (Fox Valley sharing-without-consent); no council vote, no early-exit fee, no refund. WPPA: keep the tool, write the policy.

   Why it matters

   The WI keep-path is now 7-day retain + peer-policy sharing, not a dashboard. Next door, expiration-date bagging needs no vote. Bid both in the same metro.

   Source

   WBAY
5. 5

   OpenEye OWS → CENTEGIX CrisisAlert — panic button pulls live nearby cameras

   Key point

   Cloud-to-cloud: a CENTEGIX CrisisAlert wearable activation auto-associates nearby OpenEye Web Services cameras and drops live video into the CENTEGIX Safety Platform — no app switch. School District of Manatee County, FL is among the first. OpenEye VP Systems Engineering Jon Swatzell: faster situational awareness on campus. Same week OpenEye is at GSX booth #3933 with NL visual search.

   Why it matters

   If you already sell OWS into K-12, the SKU this week is panic-to-live-video, not another NVR. Ask whether the district already has CENTEGIX badges.

   Source

   GlobeNewswire via Markets Insider
6. 6

   Universal Robots Gen 7 — AI-ready flange, CB7 +40% compute, cameras without a dress kit

   Key point

   Teradyne’s UR launched Gen 7 as a physical-AI cobot platform: new g-Series arms (UR10g-1750 / UR17g-1300 / UR18g-950), CB7 Core controller (40% more compute, 30% smaller; talks to PLCs/HMIs/MES/IPCs/vision, two form factors), PolyScope X with open APIs + ROS2 for edge-AI on an external IPC, TP7 pendant, SP7 Smart Panel at the flange. Tool flange now carries data, power, and safety so cameras and high-bandwidth sensors skip the dress kit; force-torque + impedance + RTDE onboard. PLd Cat. 3; g-Series TÜV ISO 10218-1 / UL1740; IEC 62443-4-1 ML3. IMTS demos with Cognex, Schunk, Magswitch, Smooth Robotics, Psyonic, Robotiq, Inbolt. Inbolt CEO: vision runs on the UR controller — no extra vision PC. CB7 also drives existing e-Series / UR Series arms.

   Why it matters

   This is a camera + cobot cell, not a locked dome. Spec the flange ports and CB7/IPC split so you can swap YOLO/VLM later. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.

   Source

   Universal Robots
7. 7

   arXiv: PhysBrain 1.5 — 8B VLM → physical foundation model (understand / act / predict)

   Key point

   Unified model for physical environments: language, end-effector motion, and dense visual targets as discrete sequences, trained autoregressively. Embodied pre-train from human interaction videos, then SFT on human demos + robot traj + sim. 8B scores 72.5 avg across 28 embodied-understanding benches — new open-source SOTA, on par with GPT-6-Astra and Gemini 3.6 Flash; best open-source on 14 benches; still a general multimodal model. Qualitatively emits end-effector trajectories and future scenes as spatially aligned RGB / depth / robot-mask. Project: deepcybo-physai.github.io/PhysBrain-1.5.

   Why it matters

   Plant/AMR RFPs will ask for a physical foundation model, not another locked camera brain. Keep the 8B backbone, swap the head when the site’s motion vocabulary changes.

   Source

   arXiv:2609.14973 (DeepCybo Team; Yu Bin, Haipeng Cao, et al.)
8. 8

   arXiv: BVB / Blender-VideoBench — if the agent understood the video, it can rebuild it in Blender

   Key point

   Video-understanding benches still do QA. BVB asks agents to programmatically reconstruct real-world video as animated Blender scenes via a shared Mini-BVB harness and cost cap. Scores: Dual VQA (spatiotemporal facts kept) + Latent Similarity (perceptual match). 51 configs / 10 families. Best model hits 88.6 latent similarity but keeps only 53.7% of source-correct spatiotemporal answers. Extra reasoning helps looks, not facts. 15-rater blind study: latent similarity tracks human preference. Semantic retention is the unsolved bit.

   Why it matters

   “NL search found the clip” is not the same as the model knows what happened. When a VMS vendor demos Magic Search this week at GSX, this is the test: can it reconstruct the scene, or only retrieve a thumbnail?

   Source

   arXiv:2609.15478 (Tang, Shimada, Meng, et al.)