The problem

Robotics and “physical AI” is one of the noisiest sectors to follow right now — dozens of humanoid startups, industrial incumbents, surgical robotics divisions buried inside medtech conglomerates, and a wave of ETFs slapping “robotics” on funds that are mostly holding chip and automation-equipment suppliers. Keeping an accurate, current picture of who’s actually building robots, who’s public vs. private, and what’s real signal vs. marketing takes constant manual research. That’s not a personal itch — it’s the exact desk-research problem an investment analyst, corp-dev team, or competitive-intelligence function does by hand every quarter: build a company universe, verify it against what the market actually holds, and keep it current.

I built robotics-ai-daily to turn that into a maintained system instead of a one-off spreadsheet: a database that gets researched and refreshed, feeding daily video and article output on cogniflow-ai’s robotics vertical.

The approach

What exists right now is the research substrate: a SQLite database of 59 tracked companies across 8 categories — industrial/factory automation, humanoid, logistics/warehouse, surgical/medical, consumer, drone, autonomous vehicle, and physical-AI platform (NVIDIA Isaac/GR00T, John Deere’s autonomy stack). Each row carries ticker, IR page, YouTube channel, X handle, and sourcing, split roughly 27 public / 32 private. I cross-checked that list against 9 robotics/AI ETFs (ROBO, BOTZ, IRBO, ROBT, ARKQ, UBOT, plus three European UCITS funds) to see which of my tracked companies actually show up in real fund holdings — and, more usefully, which “robotics ETF” top-10 holdings turn out to be chip and automation-equipment suppliers (Rockwell Automation, Keyence, SMC, Novanta, Yokogawa, and others) riding the label rather than companies that build robots.

The daily pipeline this feeds is designed but not yet built: pick one company or story as the day’s focus, pull the relevant YouTube clip(s) for it, research that pick’s financials and IPO signals (real filings for public names, funding rounds and valuation for private ones), generate a 15-20 minute video and a companion article, and publish both to robotics.cogniflow-ai.com — with an investing angle that draws on a separate project, canslim-analyzer, for the “how would you actually invest in this” layer. There’s no dashboard or CLI yet; the project is at the data-and-plan stage, not the running-pipeline stage.

What I learned

The most useful discipline so far was refusing to fake completeness: I only found confirmed YouTube channels for 10 of the 59 companies and X handles for 16, and left the rest null rather than guessing — a database you can trust partially beats one that looks full but is quietly wrong. The ETF cross-check was the other real finding: most companies that show up in “robotics” fund top-10 holdings aren’t robot makers at all, they’re the industrial suppliers underneath the story (sensors, motion control, chips) — a distinction that matters if you’re using ETF composition as a signal for where the sector’s money actually is.

I also hit a more mundane lesson trying to offload the full pipeline design to a background agent: the dispatch failed repeatedly on an orchestration-tooling bug (a worktree-entry timeout), not on the actual research task. Automating the research is one problem; automating the automation is its own, separately fragile layer.

Where this could go

The near-term work is finishing what the plan already scoped: daily-focus selection logic, a financial/IPO-signal research pass per company, and the video/article generation and publishing steps — plus resolving a real constraint the plan flagged and didn’t gloss over: YouTube’s ToS and copyright exposure around downloading and re-using clips from other channels, which needs a fair-use/licensing stance before any video pipeline ships.

The bigger pattern generalizes past this one sector page. Any team that tracks a competitive landscape — robotics, or any other fast-moving vertical — is doing the same manual cycle: build the company universe, verify it against what funds/filings actually show, keep it current, and turn it into something readable for stakeholders on a cadence. A maintained, sourced, gap-honest database with a repeatable publishing layer on top is a pattern an analyst or comms team could run for any sector they’re expected to stay current on, not just robotics.

Takeaway

Right now this is a well-sourced 59-company map of physical AI with the ETF noise filtered out — real research, not yet a running daily pipeline, and honest about both.