Why have people collect data instead of remotely controlling robots?

  • Robot foundation models lack vast quantities of physical interaction data that can be readily harvested. Each hour of conventional teleoperation data collection requires an available robot and an operator who is usually trained, making it costly and slow to scale.
  • Generalist has people use a handheld “Data Hand” gripper to perform everyday tasks while recording visual and motion data. No robot needs to be present, allowing data collection at multiple locations in parallel. The company says its dataset grew from about 270 thousand hours in November 2025 to more than 500 thousand hours in April 2026.
  • This approach bets that real-world interactions can preserve physical dynamics and narrow the gap between simulation and reality. Compared with approaches focused on simulation or internet video, its advantage depends on whether collection scale can continue to translate into model capability.

What do a 99% success rate and learning a new task in seconds each mean?

  • According to the company, GEN-1, released in April 2026, achieved an average success rate of 99% after post-training on individual tasks, up from 64% for the previous model. It also executed tasks about three times as fast. These results reflect performance after dedicated training.
  • GEN-1.5, released in August, changes how new tasks are taught: it can perform a task after a single 3–12-second demonstration without updating model parameters; its average success rate across 10 tasks was 59%. After adding 5 minutes of data and 10 parameter update steps per task, that figure rose to 83%.
  • Morgan Stanley values its potential to reduce the cost of teaching new tasks and eliminate days of data collection. But these results are limited to simple, short-sequence tasks; the company also acknowledges that skills learned from an immediate demonstration are more fragile than those acquired through dedicated fine-tuning.

When will more real-world data become a defensible advantage?

  • Generalist says model performance improves predictably as pretraining data increases. If this pattern continues to hold, companies will be better able to estimate returns on data collection investment and allocate computing resources, and competition will shift toward who can obtain more diverse data.
  • A defensible advantage also requires data collected by human hands to transfer to different robot bodies. Differences in arm structures, joints, and sensors can fragment data; Morgan Stanley notes that the company’s scale-effect and success-rate data are mainly self-reported, with limited independent evaluation.
  • The company hopes to secure deployments with a larger initial dataset, then strengthen its models using data fed back from those deployments. But competitors can also accumulate real-world data once deployed at scale, and the effective combination of simulation, video, and real-world data remains an open question.

Why won’t revenue immediately follow a more capable model?

  • Generalist is currently exploring deployment through hardware partners and an early access program for a small number of customers. Deliveries may include complete robots or run on customers’ existing robotic arms; the commercial path is still taking shape.
  • Morgan Stanley suggests that building complete robots in-house could strengthen control over reliability and provide deployment data, but would increase capital and operational burdens. Licensing to hardware partners is a lighter approach, but revenue is constrained by partners’ channels, and partners may switch models.
  • Manufacturing and logistics customers have long sales cycles and demanding reliability requirements, and robot failures can halt production lines. At the same time, data collection and training are capital-intensive, and pricing has yet to be set. The bank therefore believes recurring revenue ultimately also depends on robot manufacturing capability and adoption by customers in the real economy.