Robot Learning Evolution Explains Humanoid Investment Boom
Contemporary robot learning history from rules to data-driven models reveals overlooked scaling patterns and sim-to-real advances behind 2025's quadrupled humanoid investments.
Robotics has shifted from exhaustive rule-writing to simulation-based reinforcement learning and predictive foundation models.
The Technology Review piece traces early rule-based failures exemplified by Jibo's 2014 launch and 2019 shutdown despite $3.7M raised, plus the 2015 pivot to reward signals in simulation for cloth folding, yet understates the foundational 'Bitter Lesson' pattern where general computation-driven methods outperform hand-designed rules (Sutton, 2019; Technology Review, 2026).
Post-2022 transformer models adapted from LLMs ingest images, sensor data and joint states to output motor commands, as scaled in RT-X co-training across robot embodiments (Brohan et al., 2023). Coverage missed sim-to-real transfer via domain randomization and 2024-2025 real-world fleet data efforts that bridge the Jibo-era gaps.
Synthesizing these with autonomous driving history shows end-to-end learned policies consistently prevail; embodied AI breakthroughs follow identical scaling curves once interaction data volumes increase by orders of magnitude.
AXIOM: Robot learning mirrors language model scaling where general methods win; expect the next advances from amassing trillions of real-world interaction tokens via deployed robot fleets rather than new algorithms.
Sources (3)
- [1]How robots learn: A brief, contemporary history(https://www.technologyreview.com/2026/04/17/1135416/how-robots-learn-brief-contemporary-history/)
- [2]The Bitter Lesson(http://www.incompleteideas.net/IncIdeas/BitterLesson.html)
- [3]RT-X: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control(https://robotics-transformer-x.github.io/)