Imitation learning is a family of machine learning methods in which a system learns to perform a task by copying examples of an expert doing it, rather than by trial and error against a score. The expert is usually a person: a driver at the wheel, an operator steering a robot arm, a contractor writing the ideal answer to a chatbot prompt. The system records what the expert saw and what the expert did, then learns to map one onto the other. It exists because many tasks are easy to demonstrate and hard to specify. Nobody can write a reward formula for "drive politely" or "fold a shirt neatly", but a person can do both in front of a camera.

The approach sits at the base of most physical robotics work and, under a different name, inside the training of large language models (LLMs) and the AI agents built on them. Its well-documented failure modes explain why it is rarely used alone.

How learning from demonstrations works

Training starts with demonstrations. Each demonstration is a sequence of pairs: an observation, such as a camera frame, a screenshot or a text prompt, and the action the expert took in response, such as a steering angle, a set of joint positions or a written reply. A robot demonstration is typically gathered by teleoperation, in which a person drives the robot through a controller while its sensors record. A driving demonstration is a logged trip. A language demonstration is a prompt paired with a human-written answer.

The learner is a policy, a function that chooses an action given an observation. Training adjusts it until its actions match the expert's. Once deployed, it runs in a loop - observe, act, observe again - with no expert present.

Three broad techniques dominate.

Behavioural cloning is the simplest and most common. It treats imitation as ordinary supervised learning: observations are inputs, the expert's actions are labels, and the model is trained to predict the label. The term is generally traced to work on learning to fly a flight simulator from pilots' logs by Claude Sammut, Michael Bain and colleagues in the early 1990s.

Interactive imitation, best known through the DAgger algorithm (Dataset Aggregation), fixes behavioural cloning's main weakness by letting the learner drive. The policy runs, an expert labels the correct action in each state it reaches, and the labels are added to the training set before the cycle repeats. Stéphane Ross, Geoffrey Gordon and J. Andrew Bagnell of Carnegie Mellon University posted the method to arXiv on November 2, 2010.

Inverse reinforcement learning takes a step back. Rather than copy actions, it infers the reward function the expert appears to be optimising, then trains a policy against that inferred reward using reinforcement learning. Andrew Ng and Stuart Russell set out the problem in "Algorithms for inverse reinforcement learning" at the International Conference on Machine Learning (ICML) in 2000. Pieter Abbeel and Ng followed at ICML 2004 with "Apprenticeship learning via inverse reinforcement learning", tested on a simulated highway. Generative adversarial imitation learning (GAIL), posted by Jonathan Ho and Stefano Ermon on June 10, 2016, replaces the explicit reward with a discriminator that tries to tell expert behaviour from the learner's.

Origin and evolution

The first widely cited demonstration came from the road. Dean Pomerleau's ALVINN (Autonomous Land Vehicle In a Neural Network), presented at the Neural Information Processing Systems (NIPS) conference in 1988, was a three-layer neural network that took camera and laser range-finder input and output a steering direction for Carnegie Mellon's test vehicle. Its first training data were simulated road images.

Deep learning revived the idea. In April 2016, Nvidia researchers described a network that learned steering directly from human driving footage. Waymo's ChauffeurNet paper of December 2018 conceded that "30 million examples are still not enough", and added synthetic perturbations, such as drifting the car off course, to teach recovery.

Robotics took the method furthest. Google's RT-1 (Robotics Transformer), described in December 2022, was trained on a large collection of real-world robot demonstrations. In March 2023, Diffusion Policy applied image-generation techniques to robot actions, reporting a 46.9% average improvement across 12 tasks. A month later, Tony Zhao and colleagues at Stanford showed low-cost arms learning six difficult tasks "with 80-90% success, with only 10 minutes worth of demonstrations". In October 2023 the Open X-Embodiment collaboration pooled data from 22 robot types and 21 institutions, covering 527 skills.

Since then, the field has moved to vision-language-action models, which fine-tune a pretrained multimodal model on robot demonstrations. Physical Intelligence released π0 on October 31, 2024, Google DeepMind launched Gemini Robotics on March 12, 2025, and Nvidia published its GR00T N1 humanoid model on March 18, 2025.

Language models followed a parallel track. OpenAI's InstructGPT paper, posted on March 4, 2022, began by collecting "a dataset of labeler demonstrations of the desired model behavior" and used it "to fine-tune GPT-3 using supervised learning", before a second stage of reinforcement learning from human feedback (RLHF). That first stage, now called supervised fine-tuning, is behavioural cloning on text. OpenAI's Video PreTraining work of June 2022 carried the idea to computer control, training a Minecraft agent on keyboard and mouse actions inferred from online gameplay videos.

Why it matters for marketers

The agentic AI tools that operate media buying stacks inherit their behaviour from LLMs whose instruction-following was first taught through human demonstrations. Computer use agents, which click through browsers and ad platforms on screen, show the limits of pure imitation: OpenAI said on January 23, 2025 that its Computer-Using Agent combined GPT-4o's vision with "advanced reasoning through reinforcement learning". When Anthropic released computer use in public beta in October 2024, it said training Claude to count pixels accurately had been critical.

Demonstrations come from people, and that workforce is now a regulatory matter: noyb filed five GDPR complaints against Scale AI on October 27, 2026 over the tracking of workers on its Outlier platform, which noyb says has more than 900,000 freelancers paid to write prompts, rank answers and correct model output.

In ad tech, academic work on automated bidding, including Alibaba research on generative auto-bidding posted in 2024, trains models to reproduce high-performing bidding trajectories from historical logs, a form of imitation in which the expert is past campaign data. Alibaba's Bid2X, a forecasting model trained on bidding logs, raised gross merchandise volume by 4.65% in a two-month test against a model-based reinforcement learning system.

Limitations and disputes

The central weakness is distribution shift, also called compounding error. A cloned policy only ever saw states the expert visited; one small mistake puts it somewhere unfamiliar, where its next mistake is larger. Ross and Bagnell showed in 2010 that, for a task lasting T steps, a behavioural cloning policy's total error can grow with the square of T, against a linear bound for DAgger.

Causal confusion is a second problem. A policy can latch on to a feature that correlates with the expert's action without causing it. Pim de Haan, Dinesh Jayaraman and Levine described the effect in 2019 with a driving example: a model that sees the brake indicator light on the dashboard can learn to brake when the light is on, which fits the training data and fails on the road.

A third limit is the ceiling. A system that copies an expert is bounded by that expert, and inherits their mistakes. Data is also expensive: teleoperation needs a person for every robot-hour. Tesla, according to Business Insider reporting summarised by eWeek on August 26, 2025, moved Optimus training away from motion-capture suits and teleoperation towards workers wearing helmet-and-backpack rigs with five cameras. Whether video of people, without the robot's own sensor and motor data, carries enough information to train a robot remains contested.

There is also dispute over how much imitation will matter. Physical Intelligence's π*0.6, published on November 18, 2025, combined demonstrations with the robot's own experience and expert interventions through a reinforcement learning method called RECAP, which the company said "more than doubles task throughput and roughly halves the task failure rate" on its hardest tasks. On that account, demonstrations supply the starting point and experience does the rest.

What it is not

Reinforcement learning learns from rewards gathered by trial and error; imitation learning learns from examples of correct behaviour. In practice imitation supplies a starting policy and reinforcement learning refines it.

Supervised fine-tuning is the LLM industry's name for behavioural cloning applied to text, one instance of imitation learning rather than a separate idea.

Few-shot prompting places examples inside a model's input at run time without changing its weights; imitation learning changes them.

Knowledge distillation trains a small model to copy a larger model's outputs. It is imitation of a machine teacher, usually to compress it, rather than of a human expert performing a task.

Recent developments

In January 2026, Business Insider reported that Tesla was moving large-scale Optimus data collection to its Austin Gigafactory, according to TechSpot.

The newest work tries to reduce the need for human demonstrations altogether. Google DeepMind's Gemini 4 Argon, released on October 13, 2026, included a "Video-to-Sim-to-Real" pipeline in which the model watched a video of a person playing cornhole, wrote simulation code, and an Apollo-based robot then sank seven consecutive shots in an internal trial. The same company presented Genie 4 on October 27, 2026, describing simulated worlds as "safe training environments for robots and agents". As of November 2026, the frontier is less about copying a person's every move than about watching people once and learning the rest elsewhere.

Timeline

  • 1988 - Dean Pomerleau presents ALVINN at NIPS, a neural network that learns road following for Carnegie Mellon's test vehicle.
  • Early 1990s - Claude Sammut, Michael Bain and colleagues develop behavioural cloning from flight simulator logs.
  • 2000 - Andrew Ng and Stuart Russell publish "Algorithms for inverse reinforcement learning" at ICML.
  • 2004 - Pieter Abbeel and Andrew Ng publish apprenticeship learning via inverse reinforcement learning.
  • November 2, 2010 - Ross, Gordon and Bagnell post DAgger to arXiv; it appears at AISTATS in 2011.
  • April 25, 2016 - Nvidia describes end-to-end learning of steering from human driving.
  • June 10, 2016 - Ho and Ermon post generative adversarial imitation learning.
  • December 7, 2018 - Waymo posts ChauffeurNet.
  • March 4, 2022 - OpenAI posts InstructGPT, with supervised fine-tuning on labeller demonstrations.
  • June 23, 2022 - OpenAI posts Video PreTraining for Minecraft.
  • December 13, 2022 - Google posts RT-1.
  • March 7, 2023 - Diffusion Policy is posted.
  • April 23, 2023 - ACT shows bimanual tasks learned from 10 minutes of demonstrations.
  • October 13, 2023 - Open X-Embodiment pools data from 22 robot types.
  • October 31, 2024 - Physical Intelligence releases π0.
  • January 23, 2025 - OpenAI announces its Computer-Using Agent.
  • March 12, 2025 - Google DeepMind launches Gemini Robotics.
  • March 18, 2025 - Nvidia publishes GR00T N1.
  • August 2025 - Tesla's shift of Optimus training from teleoperation to camera rigs is reported.
  • November 18, 2025 - Physical Intelligence publishes π*0.6 and RECAP.
  • January 2026 - Tesla's plan to move Optimus data collection to Austin is reported.
  • October 13, 2026 - Google DeepMind releases Gemini 4 Argon with a video-to-sim-to-real pipeline.
  • October 27, 2026 - Google DeepMind presents Genie 4; noyb files five complaints against Scale AI over Outlier.

Summary

  • Who: Robotics companies such as Google DeepMind, Physical Intelligence, Nvidia and Tesla, autonomous-driving developers such as Waymo, and AI labs such as OpenAI and Anthropic, relying on human demonstrators, teleoperators and data workers such as those on Scale AI's Outlier.
  • What: A family of training methods, including behavioural cloning, DAgger and inverse reinforcement learning, in which a system learns a task from recorded examples of an expert performing it.
  • When: Demonstrated on a vehicle in 1988, formalised through inverse reinforcement learning in 2000 and DAgger in 2010, and scaled in robotics and language models from 2022.
  • Where: In the training of robot policies, driving systems, the supervised fine-tuning stage of LLMs, and computer use agents.
  • Why: Many tasks are easy to show and hard to score; copying an expert gives a system a working starting policy, though distribution shift, causal confusion and the expert's own ceiling mean it is usually combined with reinforcement learning.