IVS Briefs
IVS / Edge AI Research Brief
Overall assessment
Signal level
High on closed votes (Sheboygan unanimous, Bastrop category ban, Fox Valley cluster) and on streaming/physical-AI lab. Medium on new silicon SKUs.
Notable trends
- Two post-Flock products: Sheboygan/Bastrop = no ALPR (Bastrop writes it into policy); Kenosha = 1:1 Outpost + live video, Lightpost next.
- Sheboygan’s MSA/Traffic Analytics exhibit is what Menasha/Neenah/Waupaca chiefs cited Tuesday.
- Lab is streaming VAD + world-frame motion + sim, not another city ALPR paper.
Gaps
No new VMS SKU. No new Hailo/DeepX/Qualcomm camera-SoC in 48h (QNX/USI/Jetson reasoning already 09-08). Liberty Hill vote not closed. Stevens Point kept to honorable so this is not a 6/6 Flock brief.
Mix counts
LPR 3 / VMS 0 / Edge 1 / Policy 1 / Lab 4 (LPR+Policy = 4/9 = 44%).
Now vs On deck
Now 5 (items 1–5) · On deck 4 (ReactVAU, MINT, Antioch, AGC-VLN).
Signals (9)
- 1
Key point
Sheboygan Common Council voted unanimously Tuesday night to terminate Flock. 15 cameras were already bagged after PD/admin said Flock’s actions (Traffic Analytics sales pitch to Engineering for 5 years of speed/volume from the city’s own poles) breached the MSA and killed trust. Chief Kurt Zempel after the vote: not actively looking for a replacement, and not ignoring other tech. Public commenter Joseph Joslin urged council not to replace Flock with Axon or any other mass-surveillance vendor.
Why it matters
Closed vote, not a bag-and-wait. The Kenosha fork (keep ALPR, swap vendor) is not the Sheboygan fork. If you bid “post-Flock” here, the packet is no ALPR, not Outpost.
Source
FOX6; TMJ4 - 2
Key point
Bastrop City Council terminated Flock immediately. Councilmember Cynthia Meyer added a resolution to block future automatic license plate reader contracts, not only Flock. 12 cameras were covered last week. Mayor Ishmael Harris: the tech helped, but misuse/privacy/control concerns won. Resident John Bush wants a charter amendment so the next council cannot reverse. College Station vote Sept 24; Bryan has no date.
Why it matters
This is the first Central Texas category ban in this wave, not a vendor swap. An Axon/Motorola/Verkada LPR bid is dead on arrival unless that resolution is repealed. Copy the Meyer language into every “what does terminate mean” council memo.
Source
KBTX; KXAN - 3
Key point
Menasha council 7–1 (Tom Marshall no) to stop funding Flock on Chief Matthew Albrecht’s rec: he will not stand in front of the community and recommend Flock as the ALPR vendor after Sheboygan. Neenah Public Services and Safety Committee unanimous to stop funding; Chief Aaron Olson says PD is done with Flock, council next week is the contract vote; cameras already off Friday. Both chiefs cited Sheboygan’s resolution: Flock offered DPW anonymized volume/type/speed data in violation of the MSA. Flock to FOX 11: does not sell data, points at the Traffic Analytics blog. Waupaca County agencies jointly suspended fixed ALPRs; Law Enforcement Committee moving to cut next year’s funding ($34,500 this year). No local-misuse finding — the issue is control once data leaves via out-of-state FOIA. Albrecht/Olson still want ALPR, waiting on state/federal rules before a new vendor.
Why it matters
Sheboygan’s MSA/Traffic Analytics packet is now the exhibit other WI chiefs hand to council. Integrators: two products in one RFP — keep plates / fire Flock (Menasha/Neenah) vs county-wide suspend (Waupaca). Do not bid a national-lookup graph.
- 4
Key point
Follow-up to yesterday’s $14.6M announcement: Kenosha is replacing 31 Flock poles with 31 Axon Outpost. CEO Rick Smith (Axon Week 2025): “ALPR and live video in one unit.” Ald. Art Debaere: no local Flock misuse, moving because of optics. FOX6: Axon deal signed in January; Outpost was not named in a public meeting for eight months. Chief Patrick Patton (after six declined interview requests): other chiefs told him not to talk to FOX6; he will not buy Lightpost (streetlight that is also an ALPR) yet. FOX6 Investigators now count ≥170 officers accused of Flock-data abuse nationally (11 in WI), up from 148 last week. Privacy Bee’s Harry Maugans: “ALPR is not going anywhere.”
Why it matters
The post-Flock bid in Kenosha-class cities is a fixed LPR + live video appliance on the same poles, inside the BWC/Taser/evidence cloud. Lightpost is the 36-month refresh question: do you sell a camera or a streetlight? Nationwide lookup stays off by default — that is the privacy slide.
- 5
Key point
Hamed Tabkhi’s TeCSAR lab: learn normal movement, predict the next motion, flag departures — not a labeled weapons/fight classifier. Pose → motion tokens (SPARTA, Shopformer, PoseLift, TSGAD, ALFred). No facial recognition, no clothing/skin as features. A Jetson appliance sits on the closed IP-camera network, taps selected streams, multi-cam. NSF Civic Innovation Challenge pilots in Charlotte city center. Retail: unlabeled site drift via periodic adaptation; human is the decision-maker, AI is an attention mechanism.
Why it matters
The job is a Jetson on the existing VMS, not a camera rip-and-replace. Pose-only is the privacy slide campuses/retail will ask for this quarter. Accelerated compute beats locked “AI cameras”: swap models vs a 36–84 month hardware refresh.
Source
EE Times (Pat Brans) - 6
Key point
Offline VAU with global sampling is not causal, so it cannot run on live CCTV. Streaming VLMs compress away rare short events and burn an MLLM on hours of normal video. ReactVAU: lightweight Fast detector (Spatial Grid Folding) always on; Anomaly-Aware Persistent Memory so the cue does not decay; Slow reasoning MLLM sleeps until the fast path flags a clip, then verifies and writes a causal description. Competitive detection + reasoning under strict streaming constraints, fewer heavyweight calls.
Why it matters
This is the architecture to demand on a Jetson/Thor box: cheap always-on filter, expensive VLM as a second look. Yesterday’s WSVAD paper was event-level MIL; this one is live stream + cost.
- 7
Key point
Foundation model jointly predicts camera trajectory, FoV, camera-frame hands, and observability, then lifts to world-space hand motion. Open labeling pipeline turns public ego video into supervision; pretrain on pseudo-labels, fine-tune on a small high-quality set. HOT3D: 0.945 frame acc, 13.6 mm PA-MPJPE-p; camera 4.69 mm / 0.284° RPE; 3.67× faster than the labeling pipeline. Releases model, code, pipeline, and a 1,021-hour ego trajectory set.
Why it matters
Occupancy, robot-in-aisle, and body-worn “what did the hands do” jobs are moving from 2D boxes to world-frame tracks. Spec a swappable model on Jetson/Thor, not a locked camera SKU, when the customer asks for 3D/physical AI in the 36–84 month refresh.
- 8
Key point
NYC physical-AI sim startup (Tesla/Meta RL alumni, founded May 2025) closed $32M Series A led by Greylock; $44.75M total. Amazon Ring CPO: sims match held-out physical tests, so Ring can cut live test programs. Stack: NVIDIA Omniverse / Isaac Sim / Isaac Lab; complementary to NVIDIA Cosmos and World Labs (does not train its own world model). Domain: warehouse robots, ag, drones, fixed perception.
Why it matters
Sim-to-real is now a budget line next to the camera PO for anyone selling robots, drones, or multi-cam perception. Ring as a named customer is the security-adjacent proof. World Labs/Marble stay the world-model layer; Antioch is the eval harness.
- 9
Key point
First training-free baseline for air-ground vision-and-language navigation. UAV paints the UGV pose and VLM-anchored GOAL on a shared BEV; UGV plans road-following with a frozen VLM; UAV runs 3D-SPF in the nadir view. CARLA-Air Town10HD, 100 episodes: 77% joint success (+27 vs weaker solo agent, +24 vs Travel UAV). Naive semantic radio hurts. Code: github.com/ZSN2024/AGC-VLN.
Why it matters
Drone + ground camera jobs (campus, plant, perimeter) fail when each agent only sees its own stream. A shared BEV with labeled CAR/GOAL is the interface to copy; do not assume two VLMs talking text will cooperate.