Devices & Robotics — W33: Dyna-2 learns from human video, and Xiaomi decides it wants to build robots

Two things happened this week that matter if your AI has to touch the physical world: a robotics lab claimed it can train manipulation without teleoperation, and a consumer electronics giant with real factories decided it wants in. Everything else on this list is about where inference physically runs.

Dyna Robotics introduces Dyna-2, pre-trained on 1 million hours of human video. Dyna-2 is a world-action model trained on footage of people doing things rather than on teleoperated robot demonstrations. That is the interesting part: teleoperation costs roughly one human-hour per robot-hour of data and has been the field's hard ceiling for a decade. If human video transfers to robot action at usable fidelity, embodied AI gets a scaling curve that looks like the one language models rode. Note what's missing — a million hours is an input claim. The number that decides this is task-success on transfer, and it hasn't been published. Via Marktechpost.

Xiaomi moves on robotics. The humanoid field is mostly venture-funded specialists shipping tens of units. Xiaomi arrives with a supply chain, contract-manufacturing relationships, and retail distribution already built — which addresses the part of humanoid robotics nobody has solved. A bill of materials is not a production line, and every impressive demo so far has been hand-assembled. No product, price, or ship date surfaced this week, so file this as intent rather than a launch, but it's the first entrant whose constraint is design rather than manufacturing. Via AI Search.

LTX-2.5 puts a video world model on one workstation. Lightricks shipped LTX-2.5 with open weights and NVIDIA acceleration, positioned as a world model that generates coherent scenes rather than a clip generator. The deployment consequence is the story: video generation stops being a metered API line item and becomes a fixed hardware cost you amortise. For anyone whose compliance posture rules out sending unreleased product briefs to a hosted endpoint, this is the first credible local option — and it changes the make-or-buy maths for any team producing SKU videos or localised variants at volume. Via Marktechpost, plus a live prompting session with the LTX team on The Neuron.

Lenovo posts a 43% Q1 revenue jump on AI PC and infrastructure demand. Lenovo surged in Hong Kong on the print. This is the cleanest AI-hardware demand read-through available right now, because Lenovo sells into the mid-market — it tells you whether NPUs in laptops are actually moving units, not just whether four hyperscalers are buying GPUs. The caveat is that infrastructure and AI PCs aren't split cleanly in the headline number, so some of that 43% is rack, not desk. Via Reuters.

Anthropic confirms an in-house AI chip team. Anthropic acknowledged a silicon effort, joining Google, Amazon, OpenAI and Meta. No design, partner, or timeline is public. Custom accelerators are a five-to-seven-year bet with billions in non-recurring engineering, and they only pencil if you believe inference demand compounds indefinitely — so read this as an internal forecast stated in capex. For hardware buyers, the read-through is that the accelerator market late this decade has several more credible entrants and the single-vendor era has a visible end date. Via the Creators' AI weekly digest; unconfirmed details beyond the acknowledgement.

The tell for the robot half of this list isn't another model announcement — it's a number. Dyna needs to publish transfer fidelity on real manipulation tasks; Xiaomi needs to publish a price and a date. Until either shows up, the most deployable thing on this week's list is a video model that runs on a GPU you can already order.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.