Research
Flexion is building a reinforcement learning and sim-to-real platform for humanoid robots, aiming to address the teleoperation and data problem in robotics. The company argues that the dominant approach of using teleoperation and human demonstration at scale faces structural limitations that prevent generalization.
Teleoperation datasets are over 100,000 times smaller than what s used to train today s language and vision models. That gap doesn t close by hiring more operators, because the real world never stops changing: A shelf moves, a door handle is slightly different, and a new package type shows up on the line.
Flexion highlights that teleoperation datasets are over 100,000 times smaller than those used for language and vision models, and the real-world variability means the problem grows faster than the workforce can address. Data quality is also an issue, as operators cannot feel what they are touching or judge depth reliably, leading to slow and overcorrected movements that the robot learns from.
The company also notes the human cost of data generation, as the industry recruits workers in lower-wage economies to generate demonstrations. Flexion’s platform aims to provide an alternative by using reinforcement learning and simulation to generate training data without relying on teleoperation.
Source: The Robot Report