The daily brief on robots that learn from video

Researchers and engineers working on robot learning, imitation learning, VLAs and world models, plus people working their way into the field and the founders and investors tracking physical AI. It is written for people who want the depth: the issues are dense by design, It can take upto 15 to 20 involved minutes.

No language model searches the web, picks the stories, or writes the news. The pipeline has three stages: search pulls ~1,600 items a night from six sources, dedup collapses reposts and mirrors, and selection ranks what's left by topical relevance. A language model is used in one place, after selection: tidying arXiv abstracts and X posts so they read cleanly; titles and summaries otherwise stay in the source's own words. Every item links to the original, and each issue's footer prints the keywords its ranking optimized for, so you can audit both.

"Physical AI" is too broad to track as one query, so I split it into six micro-niches: learning from human video, latent actions and world models, video models as policies and simulators, VLA foundation models, teleoperation and data engines, and rewards and evaluation. Each niche is tracked separately, with its own sources and its own relevance profile. The run is nightly, in three stages. Sweep: the full arXiv announcement batch plus the ranked top of the other five sources. On a typical night this is about 1,600 items. Score: items are scored for topical relevance against each niche's profile. The niche results are merged, deduped, and cut to the ~30 highest-scoring. Typeset: the remaining items are laid out as a broadsheet: front page, research, newsletters, the discourse. A slow news day produces a thinner issue. The pipeline does not pad.

Each niche has an anchor: a description of the niche plus the keywords that define it. Every incoming item is scored by semantic similarity to that anchor, and each source has its own cutoff. Other factors like engagement, source authority are also considered. Nothing is sponsored and nothing pays for placement.

Six sources: arXiv, X, Reddit, Hacker News, the trade press and Substack newsletters. On one real night that was ~1,000 arXiv abstracts, ~450 Reddit threads, ~100 Hacker News stories, ~90 X posts, ~50 Substack posts and ~20 trade-press articles: about 1,700 that night; ~1,600 is typical. That isn't everything: a story that only runs somewhere else can slip through.

They're weekly and cover all of robotics; this is daily and narrow: how robots learn from video — human demonstrations, latent actions and world models (JEPA and beyond), video-generation policies, VLAs, robot data engines, and policy evaluation.

Broad robotics digests already exist; a daily brief on how robots learn from video — human demonstrations, world models, video-generation policies, VLAs, data engines and evaluation — didn't, so that's the gap this fills. Right now every reader gets the same issue. The ranking runs on a keyword profile, so editions for other niches are straightforward to add: they're planned, not built, and founding subscribers are notified first.

Two options: cancel the subscription from your PayPal account, or email siddhesh@pixelstoactions.com and I'll cancel it for you. I refund the current month on request, no questions asked.