Learning from demonstration and teleoperation

How it works

Behaviour cloning fits a mapping from observation to action on demonstrated trajectories. Because the learned policy then generates its own trajectory, it encounters its own induced distribution of observations — the problem Ross and colleagues address by collecting corrective labels in the states the learner actually reaches. For teleoperated data collection, the practical consequences are concrete: an operator’s recovery behaviour is training signal, not noise, and demonstrations collected only by an expert who never makes mistakes leave the policy with no example of recovery. Who operated, on which interface, with what latency, becomes part of the dataset description.

A concrete case

Demonstrations recorded with a smooth, low-latency haptic rig transfer poorly to a policy deployed with a laggy interface: the operator was compensating for dynamics that the deployed system does not have, and the policy learned the compensation.

Questions worth asking

• Ask how many operators produced the data and how their interfaces differed.
• Ask whether recoveries and failures were kept in the dataset or filtered out.
• Ask whether any corrective data was collected in states the policy itself reached.

Sources

• Ross, Gordon and Bagnell — A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (DAgger) ↗

  A policy trained on an expert’s demonstrations encounters, at execution time, the distribution of observations that its own actions induce; the paper’s iterative approach targets good performance under that induced distribution rather than under the expert’s.

  Boundary: A learning-theoretic result with its own assumptions and benchmark experiments. It does not establish that any particular robot policy is safe, nor how much data a given task needs.

  Locator, anchor and reuse basis

  Read at: Abstract; §1 Introduction. Inspected 2026-09-19. Original paraphrase and link to the open preprint; no figures, tables or text reproduced.

  Verify by searching the source for: good performance under the distribution of observations it induces. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know.

• Kim et al. — OpenVLA: An Open-Source Vision-Language-Action Model ↗

  A vision-language-action model couples a pretrained vision-language backbone to robot action outputs; OpenVLA is a 7B-parameter open model trained on 970k real robot demonstrations, reported by its authors to exceed a 55B closed model by 16.5 percentage points of absolute task success across 29 tasks and several embodiments.

  Boundary: Those are the authors’ own reported results on their evaluation suites and embodiments. They are not an independent replication, and success rates on 29 tasks do not describe behaviour in an unseen setting.

  Locator, anchor and reuse basis

  Read at: Abstract; §1 Introduction. Inspected 2026-09-19. Original paraphrase and link to the open preprint; reported figures attributed to the authors.

  Verify by searching the source for: outperforming closed models such as RT-2-X (55B) by 16.5%. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know.

Continue

• The loop is the system: perception, state, action, consequence
• Where the demonstrations came from, and what you are allowed to do with them
• Vision-language-action models: what the success rate means

Elsewhere on this site

• Robotics: keeping training exposure out of evaluation claims