State of Robot Learning, December 2025

State of Robot Learning, December 2025

State of Robot Learning, December 2025 State of Robot Learning — December 2025 Basically all robot learning systems today (December, 2025) are pure Behavior Cloning (BC, also called Imitation Learning) systems. Humans provide (near) optimal demonstrations of a task, and machine learning models try to imitate those actions. Formally, a policy is trained in a supervised fashion — given the robot’s state (i.e. the camera images, robot joint angles, and maybe task description text), predicts the demonstrated actions (often an action chunk , e.g. the next second of ~50hz). This doc aims to describe

saved by

• Elizabeth Qiu
• HudZah
• Blake Dee
• Shiza Charania
• Donald
• Dhruv Sheth

related reading

• Generalist - GEN-0 / Embodied Foundation Models That Scale with Physical Interactiongeneralistai.com
• Causal Video Models Are Data-Efficient Robot Policy Learners | Rhoda AIrhoda.ai
• What Do Robotics Leaderboards Tell Us About The State of Robot Learning?itcanthink.substack.com
• Ch. 21 - Imitation Learningunderactuated.mit.edu
• A VLA with Open-World Generalizationpi.website
• [2603.08546] Interactive World Simulator for Robot Policy Training and Evaluationarxiv.org
• Precise Manipulation with Efficient Online RLpi.website
• A VLA that Learns from Experiencepi.website
• pistar06.pdfpi.website
• Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsarxiv.org
• Fully autonomous robots are much closer than you think – Sergey Levinedwarkesh.com
• RLDG: Robotic Generalist Policy Distillation via Reinforcement Learningarxiv.org