Robotics Roundup: How Robot World Models Scale, a Beijing Funding Round and a Shop-Floor Humanoid

Three robotics items from 8 October 2026. The first is research: a preprint from FAIR at Meta and Mila that tries to put numbers on how robot “world models” improve as they get bigger. The other two are business news: a reported funding round for a Beijing robotics start-up, and a store humanoid that its maker plans to bring to Japanese convenience stores.

RoboJEPA: scaling laws for robot world models (preprint)

Preprint: not peer reviewed. Everything below comes from the abstract and the paper.

A world model is a system that predicts what will happen next, here how a robot’s camera view will change if the arm makes a given move. A robot can then “imagine” several possible actions and pick the one that gets closest to its goal. RoboJEPA, from researchers at FAIR at Meta, Mila and other institutions, is such a model, built on Meta’s JEPA approach, which makes predictions in an abstract internal representation rather than generating video frames [1].

The authors trained models from 22M to 8B parameters on “23 public manipulation datasets spanning 12 embodiments and 26 different action spaces” [2]. Their main finding is that the model’s “imagination error”, the error in its predicted rollouts, “follows a second-order power law in compute” [1]. In plain terms, the error falls in a predictable way as more computing power goes into training, so the researchers say they can forecast how good a larger model will be before building it. They also report that real-robot planning success improves predictably with compute [1].

On a real robot. On a single Franka robot arm, given only a goal image, the 8B model succeeded on 67% of Grasp attempts, 50% on Object Lift and 27% on Pick and Place [2]. The paper also lists results for two vision-language-action models, including π0.5 at 53% on Pick and Place. The authors stress these are “contextual references rather than directly comparable baselines”, because those models take text instructions and were trained differently [2]. So this is not a head-to-head contest.

The authors say “RoboJEPA, at 8B parameters, is the largest JEPA predictor model trained to date”, and that they “release all model checkpoints together with our training and robot deployment code” [1]. They also list limits: the image encoder was kept frozen, the fits “already sit close to data saturation”, so more varied robot data rather than just more parameters is likely needed, and the model makes no use of text [2].

Livsyn Robotics: a Series A of “more than 100 million yuan” (reported)

Reported. Beijing’s Livsyn Robotics has completed a Series A round of “more than 100 million yuan”, according to RobotHOT [3]. The investors named are Huafang Capital, Daohe Yuanqi (under Daohe Longterm), Suwen Electric Energy and “an unnamed leading strategic industrial investor” [3]. China’s PEdaily describes it as a round at the hundred-million-yuan level, with the same named investors [4].

Livsyn sells what it calls infrastructure for embodied AI: a common software layer, RUDA (Robotics Unified Device Architecture), meant to let skills and data carry over between different robot bodies. Around it sit PhiAgent, which turns videos of human hands into robot training data, RoboAgent, which carries out tasks, and LivDEX, a data-collection glove [3]. The company says optical-module insertion and removal “achieved a nearly 100% test success rate under certain conditions” (company claim; the conditions are not specified) [3].

CP AI Robotics’ C1: a store humanoid aimed at Japan (reported; rollout is a plan)

Reported. At a retail-industry event in Tokyo on 8 October, CP AI Robotics, described by News On Japan as operating primarily in Thailand, showed the C1, a humanoid designed for shops [5]. It can identify products, manage inventory, restock shelves and bag purchases, and communicate in “more than 20 languages” [5].

A plan, not a deal. According to the report, “The company plans to begin full-scale deployment of the C1 in Japanese convenience stores as early as 2027” [5]. No retailer is named. Weidong Sun of CP AI Robotics pitched it as help for staff shortages rather than a replacement for staff: “We are not aiming for completely unmanned stores” [5].

What this does not prove

• That RoboJEPA’s scaling law holds at much larger sizes. It is a preprint, the encoder was frozen, and the authors say their fits are close to data saturation [2].
• That RoboJEPA beats π0.5 or other robot models. The authors say the comparison is not like-for-like [2].
• Livsyn’s performance. The near-100% figure is the company’s own, “under certain conditions” [3].
• That C1 robots will be in Japanese convenience stores in 2027. It is a company plan, “as early as 2027”, with no named retailer [5].

The Bottom Line

RoboJEPA’s authors argue that robot world models improve predictably with compute, which, if it holds up under review, would make planning bigger robot models less of a gamble. The funding and store-robot items show money and marketing moving towards general-purpose robots, but both rest on reports and company claims, and the C1’s convenience-store future is still a plan.

Sources

1. Artem Zholus et al. (FAIR at Meta; Mila – Quebec AI Institute; and others), “RoboJEPA: Scaling Robotic Latent World Models”, arXiv:2610.10515, submitted 7 October 2026 (preprint; abstract). https://arxiv.org/abs/2610.10515
2. The same paper, full text (PDF dated 8 October 2026; methods, Table 3 real-robot results, limitations). Code: https://github.com/facebookresearch/robo_jepa ; project site: https://robojepa.github.io . https://arxiv.org/pdf/2610.10515
3. RobotHOT, “Beijing embodied AI startup Livsyn Robotics raises over 100 mln yuan in Series A”, 8 October 2026 (trade press; performance figure is the company’s claim). https://www.robothot.com/beijing-embodied-ai-startup-livsyn-robotics-raises-over-100-mln-yuan-in-series-a
4. PEdaily (投资界), report on Livsyn Robotics’ (灵生科技) Series A, October 2026 (Chinese-language trade press). https://news.pedaily.cn/202610/569884.shtml
5. News On Japan, “Humanoid Robots Set to Enter Japanese Convenience Stores”, 8 October 2026 (news report; rollout is the company’s plan). https://newsonjapan.com/article/150996.php