Nomura: Robot model breakthroughs draw closer; dexterous hands to benefit first, while factory reliability remains a hurdle

A model breakthrough is approaching. How far away is factory readiness?

Nomura expects OpenAI to release a model for physical AI between December 2026 and January 2027, providing the first catalyst. But the bank characterizes this wave of progress as approaching GPT-2: capabilities are beginning to emerge, but stable production remains some way off.
Progress has already been made in learning from a small number of demonstrations. Generalist AI reported that GEN-1.5 achieved a 59% success rate across ten tests after watching one 3–12-second demonstration, rising to 83% after a short training session with five additional minutes of data. These results come from company tests covering short tasks or simulated environments.
Reliability is what ultimately determines whether customers shift from trial budgets to operating budgets. Nomura's research shows that the best robots can operate unattended for only a few hours, while factories need a week; on-site accuracy is around 60%, versus an industry requirement of at least 98%. The bank expects closing the latter gap to take 3–5 years.

Teleoperation relies on people remotely controlling robots to collect data; it is expensive and tied to specific robot models. Nomura puts the typical cost at around 700 yuan per hour. Human first-person video costs around 300–350 yuan and remains usable across robot generations, making it better suited to expanding pretraining data.
Training methods also improve efficiency. SuperMimic first learns how the body moves, then trains a controller within the model. The paper reports that training on a single GPU for 3 hours outperformed a baseline method trained on 128 GPUs for 7 days, reducing compute requirements by more than a thousandfold. This improvement concerns whole-body motion control; Nomura consequently believes that this stage may become less dependent on computing power.
Video can address initial learning, but continued progress also relies on data generated while robots work. Nomura believes that only a sufficient number of deployments can create a cycle in which operations generate data, models improve and deployments increase.

Nomura favors large AI labs and Tesla: the former have general-purpose models, talent and computing power, while the latter links simulation, data platforms, factory testing and its own robots. The two sets of capabilities can complement each other.
On-site scale remains a constraint. According to Nomura's research, Tesla aims to accumulate 1 million hours of usable data by the end of 2026 using around 10 thousand robots for internal training. By comparison, China's seven largest express delivery companies are expected to purchase fewer than 500 humanoid robots in total in 2026. The bank therefore believes the feedback loop will begin within companies that control both models and robots.

Because training data was insufficient in the past, robots more often used simple grippers. Nomura believes stronger foundation models can transfer skills learned by two- or three-finger grippers to five-finger hands, supporting greater adoption and lower unit prices.
The bank expects shipments of dexterous hands to grow 80%–120% year on year in 2026, driven mainly by high-degree-of-freedom products with more movable joints. The rationale is that models lower the barrier to using complex hands, while shipment growth is accompanied by falling prices.

Nomura expects robots to converge on a few standard forms once the model approach stabilizes. The key then will be ensuring that the more than 20 joints on each robot perform similarly: if differences between bodies are too large, a control strategy learned by one machine will be hard to reuse on others, hampering data sharing.
The bank therefore favors Unitree, automakers and key component suppliers that iterate quickly and control costs well. It also expects Unitree to increase the proportion it sources externally as volume rises, balancing cost and quality. Conversely, complete robot makers with poor product consistency that depend on customers collecting data may face declining demand.