EXPLAINER

Imitation Learning (Behaviour Cloning and DAgger)

techniquefreshness: evolvingsources not yet verified

Imitation learning trains a policy from expert demonstrations — for example human teleoperation of a robot — instead of from a hand-designed reward.

Short answer: Behaviour cloning treats demonstrations as supervised data; DAgger and related methods address compounding errors by querying the expert on the learner's own states. Modern robot policies (ACT, Diffusion Policy, VLAs) are largely imitation-learned. Data quality and coverage dominate results.

What is it?

Instead of describing the task with a reward and letting a robot figure it out, you show it how: a person guides the robot through the task many times and the robot learns to copy. It is often the fastest route to a working skill.

Also known as: behavior cloning, behaviour cloning, learning from demonstration, DAgger, teleoperation data, ACT policy, ALOHA, diffusion policy.

Technical definition

Behaviour cloning minimises the difference between policy and demonstrated actions, but suffers covariate shift: small errors lead to states not seen in demonstrations (Ross et al., 2011, DAgger). Remedies: DAgger-style interactive data collection, data augmentation, action chunking (ACT; Zhao et al., 2023), and expressive policy classes such as diffusion policies (Chi et al., 2023) that model multimodal action distributions. Inverse RL infers rewards from demonstrations; offline RL can improve beyond demonstrations. Demonstration quality, diversity and consistency are critical.

Entity type: technique · Freshness class: evolving.

When should I use it?

Default approach for manipulation skills with available teleoperation.

Implementation patterns

  • Collect consistent teleoperated demonstrations with varied initial conditions.
  • Action chunking and temporal ensembling.
  • DAgger-style corrections on failures.
  • Augmentations for camera and lighting variation.

Technologies and frameworks

Commonly associated: ACT / ALOHA, Diffusion Policy, LeRobot, robomimic. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Define task and success check.
  2. Collect 50–200 clean demonstrations as a starting range to test.
  3. Train a baseline (ACT/diffusion).
  4. Evaluate across many trials.
  5. Add corrective data for failures.

Common failure modes

  • Compounding errors off-distribution.
  • Inconsistent or multimodal demonstrations confusing the policy.
  • Too few diverse demonstrations.
  • Overfitting to a scene.

Troubleshooting

  • Robot drifts then fails — Add recovery/correction demonstrations (DAgger-style).
  • Policy averages between modes — Use expressive policies (diffusion/chunked) and cleaner demonstrations.

Contrasted with

  • Reinforcement Learning — RL learns from reward; imitation learns from demonstrations.
  • Supervised Learning — Behaviour cloning is supervised learning applied to demonstrated state→action pairs, with extra distribution-shift problems.

Questions this page answers

  • What is imitation learning?
  • What is behavior cloning?
  • What is DAgger?
  • What is a diffusion policy?
  • How many demonstrations do I need?

Typical goals: teach a robot a skill by demonstration; handle compounding errors.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: evolving — Established techniques whose best practice shifts as research and tooling move. Re-review every 6 months.

Canonical sources (unverified leads):

  • Ross, Gordon & Bagnell, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (DAgger) (2011) — arXiv:1011.0686
  • Zhao et al., Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA) (2023) — arXiv:2304.13705
  • Chi et al., Diffusion Policy (2023) — arXiv:2303.04137