Source note

Robots learn: A brief, contemporary history

Robot Learning HistoryVision Language ActionSim2realRobot Foundation ModelsHumanoid Robots

This article gives a short history of how robot learning shifted from hand-written rules to data-driven models, simulation training, and foundation-model-style policies. It argues that this shift is a main reason robotics investment and deployment rose sharply in the mid-2020s.

  • Traditional robotics depended on engineers specifying rules for each situation, which breaks down for messy real-world tasks like folding clothes, handling novel objects, or interacting with people.
  • Policies trained only in simulation often fail on real robots because small differences in friction, lighting, materials, and sensing can change outcomes.
  • General-purpose robots need to map language, vision, and state into actions across many tasks, but older scripted systems had weak language ability and poor adaptability.
  • The article traces three main learning shifts: rule-based control, reinforcement learning in simulation, and large-scale action prediction from multimodal data.
  • In simulation-based learning, robots improve by trial and error with reward signals. Domain randomization varies physics and visuals across many simulated worlds so the policy transfers better to reality.
  • In foundation-model-style robotics, systems take camera views, sensor readings, joint states, and language instructions, then predict the next robot action many times per second.
  • Some companies collect data from deployed robots in warehouses or other work settings, then use that real-world feedback to improve the model over time.
  • The piece uses case studies: Jibo for social interaction limits, OpenAI Dactyl for sim-to-real dexterous manipulation, Google RT-1/RT-2 for vision-language-action control, Covariant RFM-1 for warehouse picking, and Agility Digit for humanoid deployment.
  • Humanoid robot investment reached $6.1 billion in 2025, about 4x the amount invested in 2024, according to the article.
  • Google collected data for 17 months across 700 tasks to build RT-1. RT-1 achieved 97% success on tasks it had seen before and 76% on unseen instructions.
  • OpenAI's Dactyl later applied sim-based techniques to Rubik's Cube solving, with 60% success overall and 20% on particularly hard scrambles.
  • Covariant deployed warehouse robot systems at customer sites such as Crate & Barrel and released RFM-1 in 2024, but the article gives no benchmark table or aggregate success rate.
  • Agility's Digit is described as one of the first humanoids used for real warehouse work by Amazon, Toyota, and GXO. A concrete limitation is payload: Digit can lift 35 pounds.
  • The article is a journalistic overview, not a research paper, so quantitative evidence is selective and there is no unified experimental comparison across methods or systems.