AiWo at BMVC 2026: Why recognizing new assembly actions remains hard
We’re delighted to share that a paper by AiWo’s doctoral researcher Changyi Li has been accepted to the 37th British Machine Vision Conference (BMVC 2026), one of the major international conferences on computer vision, organised by the British Machine Vision Association.
- Conference: https://bmvc2026.bmva.org/
- Date: 23 Nov 2026 – 26 Nov 2026, Lancaster, UK
About the research
Assembly actions combine a manipulation (verb) with a part or tool (noun), so the number of possible actions grows multiplicatively while training data covers only a fraction of them. The common solution, verb–noun decomposition, predicts each component separately and recombines them, making unseen combinations reachable. This paper asks whether reachable also means recognizable, through a systematic analysis on three assembly datasets (MECCANO, HA-ViD and IMPACT).
Accepted paper and key contributions
- Reachability Is Not Generalization: Understanding Verb–Noun Decomposition in Assembly Action Recognition
Shows that decomposition generalizes only partially: accuracy on unseen actions closely follows how well their components are covered in training, and errors concentrate on the component with the larger vocabulary, making the bottleneck predictable before training. It further finds that the widely used shared-encoder design entangles verbs and nouns, and that training them independently improves unseen-action performance by up to 6×.
Why it matters
Industrial systems for assembly monitoring, robot learning and procedural guidance will inevitably meet action combinations never seen in training. The findings give practical guidance on where to invest data and model capacity, and the released diagnostics and IMPACT compositional split support future work in this direction.
Authors
Changyi Li, Yu Xiao (Aalto University)