Successful demonstrations can still make a narrow training set. Researchers at KAIST and Holiday Robotics investigate the variety between those successes in DiscoDemo, a preprint posted on 23 October.[1]
The underlying difficulty is older than today’s robot foundation models. The 2011 DAgger paper explains that an imitation learner changes the situations it encounters through its own actions. A small mistake can therefore put it somewhere the expert demonstrations never visited. Its proposed remedy gathers examples from situations reached by the learner, asks an expert what to do there, and combines those examples with the existing training data. The training set gradually becomes more relevant to the behaviour of the machine being taught.[2]
Give the teacher more than one solution
Starting from 20 human demonstrations per task, DiscoDemo’s reinforcement-learning generator earns rewards in simulation for task completion and diverse solutions. Successful, filtered examples train a separate pi0.5 imitation policy. Each generated dataset contains 3,000 demonstrations; the matched P-RFCL baseline uses the same generator with its diversity objective disabled.[1]
Real-robot success in four tasks
| Task | DiscoDemo | Without diversity reward |
|---|---|---|
| Banana into bowl | 90% | 45%Source 1 |
| Cube stacking | 75% | 35%Source 1 |
| Round peg insertion | 80% | 45%Source 1 |
| Square–circle peg insertion | 65% | 30%Source 1 |
Testing used simple tabletop scenes. The method requires accurate simulation, defined success signals and task-specific safety checks.[1]
Data volume is only one design choice
Generating extra demonstrations has an established research history. MimicGen, presented at the 2023 Conference on Robot Learning, adapted human demonstrations to different contexts. Its authors generated more than 50,000 demonstrations across 18 tasks from about 200 human examples, varying scene arrangements, objects and robot arms. Those datasets supported learning tasks such as coffee preparation and multipart assembly. It showed that a small collection of human actions could become a much larger teaching resource through systematic transformation.[3]
There are several distinct questions to ask of such a resource. How many examples does it contain? How many different starting situations do they cover? How many successful strategies remain available once a task is under way? Those questions describe different kinds of variety. A large file count answers only the first.
For a team commissioning robot training data, a useful evaluation would hold the learner and dataset size steady, then vary the experiences represented. That makes the purchase decision more informative: whether to collect more repetitions, broaden the starting conditions, or seek additional ways of completing the task. Counting demonstrations is easy. Deciding what another demonstration should teach is the more interesting engineering problem.
Sources & further reading
- DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning arXiv, 2610.32093v1 · 23 Oct 2026 · Accessed 28 Oct 2026
- Ross, Gordon and Bagnell: A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning Proceedings of Machine Learning Research, 15, 2011 · Accessed 28 Oct 2026
- Mandlekar et al.: MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations Proceedings of Machine Learning Research, 229, 2023 · Accessed 28 Oct 2026