Anu ships six releases to cut a robot station deploy to days
Embodied AI's bragging rights are moving from leaderboards to factory floors. One Chengdu company just bundled a world model, a small on-site model and a deployment system into a single pitch — and published the unglamorous numbers that decide whether any of it is a business.
Anu Intelligence (安努智能) has released six results at once, and the one that matters most is ActiWorld — a world model built to check what a robot's actions actually do rather than what the next frame looks like. Developed with HKUST (Guangzhou)'s SC2I lab and still under review, ActiWorld adds three training-only constraints to an action-conditioned diffusion backbone: recovering the commanded action from adjacent predicted frames, modelling both how the world arrived and how it evolves, and concentrating prediction error on the contact region instead of the static background. The auxiliary objectives are discarded at inference, so the middleware costs a robot nothing at run time and doesn't touch the policy layer. It beats IRASim on PSNR and latent L2 across RT-1, BridgeData V2 and Language Table, keeps generated policy rankings within 3.25 percentage points of the simulator's across four policies, and lifts LIBERO-plus success from 88.1% to 93.5%.
The real-robot number is the one worth reading twice. On a package-sorting task — QR code facing up, onto a conveyor — a fixed ACoT policy succeeded 56% of the time across 50 paired physical rollouts on an AgiBot G2. With ActiWorld scoring candidate action chunks by their imagined outcomes first, that rose to 62%: six points of absolute success, roughly a 14% cut in the failure rate. The companion planner, CA3C, only commits to a long chunk while independent predictions agree, cutting the segment short the moment futures diverge or the robot enters a contact, grasp or place. That is a modest gain from a genuinely useful idea: a robot that can cheaply imagine which grip is about to fail doesn't need a better policy to do better work.
The other five pieces close the loop from that prediction to a running production line. ActiWorld screens which attempted actions worked, filtered samples feed the RoHIL/EvoHIL reinforcement-learning system, the result is distilled into AnuVerse-0.5 — a small vertical model tuned for roughly hundred-millisecond responses so a robot can replan inside the line's cycle time — and Nexus deploys it across Agibot's bodies and other vendors' hardware, which is what keeps a customer's next station from being a fresh integration project. The training mix is real line data from Fulin Precision's Mianyang plant and Intel, plus long-tail synthetic data from its SimLab simulator and Helios video generator.
The numbers Anu attaches to that loop are the actual story. At Fulin, the first station took about four months; the same class of station now deploys in roughly a week, with three days claimed for mature, standardised stations whose bodies and interfaces are already adapted — new non-standard tasks still get re-evaluated from scratch. Engineers have been on site eight months and mean time between failures is past 300 hours. Manual interventions fell from more than ten a day to two or three a week. One robot on one shift does about 70% of a human worker's output; with battery swapping across two shifts that's roughly 1.4 workers, and Anu's own cost model puts payback at about three years. It has also wired the stack into SAP: work orders from Intel's Chengdu warehouse system now parse into robot tasks and return execution status, though that site is still in verification.
Read the caveats as the interesting part, not the fine print. Even with 95% synthetic data, the Fulin line still broke on changed lighting, randomly stacked parts, wrinkled packaging and QR codes on flexible surfaces — the long tail only shows up after months on a real line. A 70%-of-one-worker robot with a three-year payback is a foothold, not a revolution, and it is being sold into factories precisely because it is cheaper than the alternative, not because it is better than a person. Chairman Wen Hongjie's own test for whether this scales is the right one: the company has truly entered large-scale operation when the software revenue from adding one more robot grows faster than the delivery headcount needed to install it. We went long on the research side of this shift this morning in Deep Dive — Robots are learning what their actions do, not what to do.
What to watch: whether the SAP work-order pilot at Intel Chengdu converts from a verified link into a paid multi-site rollout — that's the first real test of whether Anu is selling software or just shipping projects.
Would you buy a robot that does 70% of a shift if the deployment took three days — or does the three-year payback kill it? Tell us in the comments.
