Programming a robot to fold a shirt or load a dishwasher with explicit rules is nearly impossible: the variations are endless. Instead, many of the most impressive recent robot demonstrations rely on imitation learning — showing the robot how to do a task many times and training a neural network to reproduce it.

The basic idea

In imitation learning (also called learning from demonstration or behaviour cloning), people perform a task while the robot records what it sees (camera images, joint positions) and what actions were taken. A model is trained to predict the right action from the current observation. Once trained, the robot runs the model in a loop: observe, predict action, act, repeat — many times per second.

Collecting demonstrations

The most common method is teleoperation: a person controls the robot directly, often with a second “leader” arm whose movements the robot copies, or with a VR headset and hand tracking. This produces data in exactly the robot’s own body and sensors.

In 2023, researchers at Stanford introduced ALOHA, a low-cost bimanual teleoperation system, together with a learning method called Action Chunking with Transformers. With about 50 demonstrations per task, robots learned fine-grained skills such as opening a cup lid or inserting a battery — showing that capable manipulation was possible with affordable hardware and modest data.

Better policies

Early behaviour cloning struggled when demonstrations varied: averaging different ways of doing a task produced actions that matched none of them. Newer methods handle this:

  • Action chunking predicts sequences of future actions instead of one step at a time, giving smoother, more consistent motion.
  • Diffusion Policy (2023), from researchers at Columbia, Toyota Research Institute and MIT, generates actions with the same denoising approach used by image generators, capturing multiple valid ways of performing a task and performing strongly across many benchmarks.

Scaling up data

A single lab’s demonstrations cover only a few tasks and one robot type. The Open X-Embodiment project (2023), a collaboration of many institutions, pooled data from 22 robot types and showed that models trained on this mixed data could transfer skills across robots better than models trained on each robot’s data alone. This idea underpins robot foundation models and the vision-language-action models that follow natural-language instructions.

Limitations

  • Data cost: demonstrations require people and robots, which is slow and expensive compared with web text.
  • Distribution shift: small errors put the robot in situations it never saw in demonstrations, where it may fail.
  • Generalization: new objects, lighting or rooms can break learned skills.
  • Safety and reliability: a policy that works 90% of the time is impressive in research and unacceptable in many real deployments.

Researchers combine imitation with reinforcement learning, simulation, data from human videos and corrective demonstrations to address these problems.

Why it matters

Imitation learning is one of the main reasons robots — including humanoid robots — are suddenly learning dexterous household and factory tasks. Its progress depends on data more than algorithms, which is why companies are building large teleoperation fleets.

Imitation versus reinforcement learning

Reinforcement learning lets a robot discover behaviour through trial and error and rewards, which is powerful in simulation but slow and risky on physical hardware. Imitation learning starts from human examples, which is faster to get working but limited by the quality and coverage of the demonstrations. Many systems now pre-train on demonstrations and refine with reinforcement learning.

What to watch next

  • Large-scale teleoperation data collection by robotics companies.
  • Learning from human videos to reduce robot data needs.
  • Robot foundation models shared across hardware types.
  • Reliability benchmarks for real-world deployment.

Key takeaways

  • Imitation learning trains robots to reproduce human demonstrations.
  • Teleoperation produces demonstrations in the robot’s own body and sensors.
  • ALOHA showed fine manipulation from about 50 demonstrations with low-cost hardware.
  • Action chunking and diffusion policies handle varied demonstrations better.
  • Pooled datasets like Open X-Embodiment help skills transfer across robots.

Frequently asked questions

How many demonstrations does a robot need?

For a single task in research settings, tens to hundreds; general-purpose skills require far larger and more diverse datasets.

Is imitation learning the same as reinforcement learning?

No. Imitation learning copies demonstrations; reinforcement learning learns from rewards through trial and error. They are often combined.

Can robots learn from YouTube videos?

Researchers are working on it, but human videos lack the robot’s exact actions, so they are used alongside robot data.