What changes are large models bringing to robotics?

  • A new generation of frontier models is moving beyond adapting to changes in object color and texture toward completing new tasks with zero or few examples. In unfamiliar environments and unstructured settings, reliance on task-by-task fine-tuning could decline.
  • Goldman Sachs believes this will accelerate commercialization in areas such as home assistance and complex logistics, rather than make the industry mature overnight. According to the experts interviewed, general-purpose robotic manipulation is still at an early stage comparable to the early development of autonomous driving when measured by safety and mass-production reliability, but the pace of technological progress is expected to reach 2 to 4 times the historical pace of autonomous driving.
  • There remains a trade-off between generalization and precision: industrial manufacturing demands exceptionally high reliability and repeatability and will continue to rely on dedicated data and deterministic motion control; large models are better suited to commercial, logistics and home-service environments that require flexible responses to change.

What steps remain between understanding a task and completing an action?

  • Current three-dimensional perception is still not accurate enough to handle millimeter-scale interactions reliably, and inference latency also struggles to meet the demands of dynamic stability, high-speed operation and rapid recovery. Commercial deployment also requires compressing large-model capabilities into more efficient edge models to reduce operating costs.
  • Experts divide robot intelligence into three layers: high-level models understand and plan tasks, the middle layer turns intentions into motion trajectories, and the low-level layer handles high-speed control and force feedback. Frontier model labs will find it difficult to delve deeply into the low-level layer in the near term because each machine has different electromechanical characteristics and must also undergo physical testing, wear and repair.
  • Goldman Sachs therefore believes robot manufacturers’ long-term differentiation will increasingly come from real-time control, safety architecture, edge-model deployment and hardware-software integration. Even when a third-party large model is used as the “brain,” manufacturers skilled in precision electromechanics and motion control can still handle physical execution.

Why might data collection and tactile sensing benefit first?

  • Dexterous manipulation involves multiple degrees of freedom of movement, continuous contact and tactile feedback. Experts estimate that reaching the threshold for general-purpose production may require approximately 10 million hours of high-quality multimodal interaction data, which ordinary passive or noisy video recordings can scarcely replace.
  • First-person video helps with understanding tasks but lacks information such as contact force; simulation can quickly generate motion trajectories, but details such as friction and material deformation differ from reality and require calibration.
  • Experts are optimistic about general-purpose manipulation interfaces that collect motion data directly at points of contact between the hand and an object, combined with high-resolution tactile sensing to fill visual blind spots caused by close-range occlusion and depth errors. On this basis, Goldman Sachs is optimistic about near-term demand for data-collection hardware such as three-dimensional spatial sensors and tactile arrays.

How do industry opportunities map to companies and applications?

  • Goldman Sachs believes combining foundation models with collaborative robots and existing factory automation could also be a practical solution. Whether manufacturers benefit depends on whether their sensing, data collection, motion control or key components can be incorporated into real-world applications.
  • Goldman Sachs rates Inovance Technology a buy, citing its industrial automation and motion control capabilities and its portfolio of hardware with sensing and data-collection capabilities. It also rates Shuanghuan Driveline a buy and forecasts its AI robot revenue will reach 400 million to 500 million yuan in 2027, more than doubling from its previous level.
  • Sanhua Intelligent Controls’ robotics opportunity lies mainly in actuator assembly. Goldman Sachs rates its H shares a buy and its A shares neutral, preferring the lower-valued H shares; it also believes market expectations for humanoid robots are optimistic relative to Optimus production progress.