Project AURIONAI Use-Cases Research & Innovation

Imitation Learning from Demonstrations

Training robot or game-playing policies directly from recorded human demonstrations, without a language model backbone.

Tier Grades

feasible

Tier A: Frontier AI labs

feasible

NVIDIA scales behaviour cloning to 40,000 hours of gameplay video in NitroGen.

feasible

Tier B: Indonesian AI labs

feasible

ACT learns real tasks from about 10 minutes of demonstrations on low-cost hardware, which fits a modest lab budget.

feasible

Tier C: Institution units such as ITB

feasible

ACT learns real tasks from about 10 minutes of demonstrations on low-cost hardware, which fits a university robotics lab.

feasible

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

Ten minutes of demonstrations per task on low-cost hardware fits a site proof of concept for a repetitive manipulation task.

Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

  • ComputeACT runs on a low-cost bimanual hardware setup; Diffusion Policy and ACT training compute not stated in the abstracts; NitroGen scales behaviour cloning to 40,000 hours of video.
  • DataACT: 6 real-world tasks learned at 80 to 90% success from about 10 minutes of demonstrations each; NitroGen: 40,000 hours of gameplay video across more than 1,000 games.
  • Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 4 authors on the ACT paper (arXiv 2304.13705); 8 on Diffusion Policy.
  • Time horizonMonths.
  • Shelf-life signalsDiffusion Policy reports a 46.9% average improvement over prior methods across 12 tasks, then OpenVLA reports beating Diffusion Policy by 20.4% after fine-tuning.
  • Pipeline roleTier A scales demonstration data; Tier B and Tier C reproduce ACT and Diffusion Policy on local arms; Tier D uses them for single-task proofs of concept.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars
Stanford UniversityACT (Action Chunking with Transformers)Columbia UniversityDiffusion Policy
Observed actors
Stanford UniversityStanford UniversityColumbia UniversityColumbia UniversityNew York UniversityNew York UniversityNVIDIANVIDIATsinghua UniversityTsinghua University
Linked models
Stanford UniversityACT (Action Chunking with Transformers)robot policyexemplarColumbia UniversityDiffusion Policyrobot policyexemplarNVIDIANitroGenrobot policyproduced byNew York UniversityDobb-Erobot policyproduced by

Papers

Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

  1. arxiv.org/abs/2304.13705 https://arxiv.org/abs/2304.13705
  2. arxiv.org/abs/2303.04137 https://arxiv.org/abs/2303.04137
  3. arxiv.org/abs/2601.02427 https://arxiv.org/abs/2601.02427
  4. OpenVLA: An Open-Source Vision-Language-Action Model arxiv.org https://arxiv.org/abs/2406.09246

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Imitation Learning from Demonstrations", https://aurion.semai.id/topics/imitation-learning-from-demonstrations/, records of October 2026.